33 17.00-17.30: Despina Papadopoulou, Theodore Marinis, Leah Roberts: Lexical effects in sentence processing: evidence from modifier attachment in Greek and English.. Eiter, Radach, Inho
Trang 19th Annual Conference on Architectures and Mechanisms for
Language Processing
August 25–27, 2003 Glasgow, Scotland
Organising Committee:
Simon Garrod Linda Moxey Tony Sanford Sara Sereno Patrick Sturt
AMLaP-2003
9th Annual Conference on Architectures and Mechanisms
for Language Processing
August 25–27, 2003 Glasgow, Scotland
Sponsored by the British Academy
Organising Committee:
Simon GarrodLinda MoxeyTony SanfordSara SerenoPatrick Sturt
Programme Committee:
Gerry AltmannKathryn BockMarc BrysbaertManuel CarreirasChuck CliftonMartin CorleyMatt CrockerEugene DawydiakJanet FodorFernanda Ferreira
Angela FriedericiSimon GarrodTed GibsonBarbara HemforthYuki KamideFrank KellerGerard KempenLars KoniecznyVincenzo LombardoMaryellen McDonald
Brian McElreeDon MitchellLinda MoxeyWayne MurrayNeal PearlmutterMartin PickeringJoel PynteKeith RaynerTony SanfordChristoph Scheepers
Sara SerenoSuzanne StevensonPatrick SturtMike TanenhausMatt TraxlerJohn TrueswellJos Van Berkum
Conference Helpers:
Mois´es BetancortJason BohanEugene Dawydiak
Heather FergusonScott GrahamBarbara Howarth
Tracy MacLeod
Jo MolleLorna Morrow
Peter Ward
Trang 2Monday August 25th
9.00-10.45: Registration
10.45-11.00: Opening Remarks
Chair : Patrick Sturt (University of Glasgow)
11.00-11.30: Brian McElree, Martin Pickering, Matthew Traxler, Steven Frisson: Enriched
compositional processing 1
11.30-12.00: Manabu Arai, Roger van Gompel, Jamie Pearson, Vera Schumacher: The
rep-resentation of transitivity information 2
12.00-12.30: Laurie Stowe, Monika Zempleni, John Hoeks: Processing Idioms in the Left
and Right Hemispheres 3
12.30-14.30: Lunch and Poster Session 1
Chair : John Henderson (Michigan State University)
14.30-15.00: Irina Sekerina: Grammatical gender and mapping of referential expressions in
Russian 4
15.00-15.30: Gerry Altmann, Silvia Gennari, Louise Shackleton: Situating language in the
visual world: Mapping the semantics of language onto the semantics of visual scenes 5
15.30-16.00: Elsi Kaiser, John Trueswell: Dividing up referential labor: Finnish pronouns
and demonstratives in on-line processing 6
16.00-16.30: Coffee Break
Chair : Martin Pickering (University of Edinburgh)
16.30-17.00: Jennifer Arnold, Zenzi Griffin: The role of competition in the production of
referring expressions 8
17.00-18.00 (Invited): Michael Tanenhaus: Referential Domains in Language Processing 9
19.00: Civic Reception at the Glasgow Gallery of Modern Art
Tuesday August 26th
Chair : Chuck Clifton (University of Massachusetts at Amherst)
9.00-9.30: Pia Kn¨oferle, Matthew Crocker, Christoph Scheepers, Martin Pickering: The
in-teraction of mental representations from linguistic and visual input: A processing account
of role-assignment in visual worlds 10
9.30-10.00: Fernanda Ferreira, Janet Fodor, Ellen Lau: Structure-Building and Lexical
Ac-cess During Garden-Path Recovery 12
10.00-10.30: Markus Bader, Josef Bayer: Diagnosis and Garden-Path Recovery-Case and
Agreement Symptoms Revisited 13
10.30-11.00: Coffee Break Chair : Tony Sanford (University of Glasgow)
11.00-11.30: Wind Cowles, Maria Polinsky, Marta Kutas, Robert Kluender: Brain responses
to differences in the processing of informational and contrastive focus 15
11.30-12.00: Marieke van Herten, Herman Kolk, Dorothee Chwilla: How semantic analysis
can overrule syntactic analysis: an ERP study 17
12.00-12.30: Albert Kim, Janice Chen, Caitlin Rippey, Lee Osterhout: Combinatory
seman-tic processing can occur independently of syntacseman-tic support 18
12.30-14.30: Lunch and Poster Session 2 Chair : Christoph Scheepers (University of Dundee)
14.30-15.00: Zenzi Griffin: Precision Timing in Speaking 19
15.00-15.30: Robert Hartsuiker, Martin Pickering, Nivja de Jong: Semantic facilitation and
phonological interference in self-correction: evidence from picture naming 20
15.30-16.00: Victor Ferreira, Robert Slevc, Erin Rogers: How do speakers avoid linguistic
ambiguity? 21
16.00-16.30: Coffee Break Chair : Simon Garrod (University of Glasgow)
16.30-17.00: Sarah Brown-Schmidt, Michael Tanenhaus: Real-time circumscription of
ref-erential domains in non-scripted task-based dialog 22
17.00-18.00 (Invited): Jonathan Ginzburg: What clarification requests tell us about dialogue
information states 23
19.00: Conference Dinner and Ceilidh at College Club
Trang 3Wednesday August 27th
Chair : Matt Crocker (Universit¨at des Saarlandes)
9.00-9.30: Luca Onnis, Padraic Monaghan, Nick Chater, Korin Richmond: Phonology
im-pacts segmentation and generalisation in speech processing 24
9.30-10.00: Morten Christiansen, Florencia Reali, Padraic Monaghan, Nick Chater:
Multiple-Cue Integration in Language Acquisition: The Differential Contribution of
Phonological and Distributional Cues 25
10.00-10.30: Frank Keller: A Psychophysical Law for Linguistic Judgments 26
10.30-11.00: Coffee Break
Chair : Linda Moxey (University of Glasgow)
11.00-11.30: Julia Simner, Martin Pickering: Anticipating Cause and Consequence during
Text Comprehension 27
11.30-12.00: Martin Corley, Lucy MacGregor, Bryony Tilley: Task demands and the
alloca-tion of attenalloca-tion in reading: some early investigaalloca-tions 28
12.00-13.00 (Invited): Peter Hagoort: How the brain solves the binding problem for language 29
13.00-14.30: Lunch and Poster Session 3
Chair : Julia Simner (University of Edinburgh)
14.30-15.00: Roger van Gompel, Asifa Majid: Accessing antecedents: Pronouns with
infre-quent antecedents are easier to process than pronouns with freinfre-quent antecedents 30
15.00-15.30: Raymond Bertram, Jukka Hy ¨on¨a, Alexander Pollatsek: Compound processing
in Finnish: a survey 31
15.30-16.00: Jeremy Pacht, Walter Schroyens, G ´ery d’Ydewalle: Lexical competition in
vi-sual word recognition: Evidence from foveal and parafoveal form-priming paradigms 32
16.00-16.30: Coffee Break
Chair : Don Mitchell (University of Exeter)
16.30-17.00: Eva Fern ´andez, Dianne Bradley, Jos´e Manuel Igoa, Celia Teira: Prosodic
phrasing in the RC-attachment ambiguity: Effects of language, RC- length, and position 33
17.00-17.30: Despina Papadopoulou, Theodore Marinis, Leah Roberts: Lexical effects in
sentence processing: evidence from modifier attachment in Greek and English 35
17.30-18.00: Shravan Vasishth: Processing noncanonical constructions in free word order
languages 36
Poster Session 1, Monday August 25th
1 Keller: Evaluating Probabilistic Models of Human Parsing 39
2 Hemforth, Konieczny, Pynte: Event-types in pronoun resolution 41
3 Konieczny, Schimke, Hemforth: The dynamics of plural-markings in agreement errors in
production 42
4 Pynte, Portes, Holcomb, Di Cristo: Relative Clause Attachment in French: an ERP study 43
5 Scheepers, Klee: Pre-verb thematic role assignment and -revision in German verb-final
constructions: Counter-evidence from eye-movements in reading 44
6 O’Bryan, Folli, Harley, Bever: Event structure effects on garden pathing 45
7 Harrison, Branigan, Hartsuiker, Pickering: Three way attraction effects in Slovenian 46
8 McCauley, Shillcock: Does the Asymmetric Neighbourhood Effect interact with
handed-ness in English Readers? 48
9 Sanford, Bohan, Sanford, Molle: Detecting text changes as a function of load and extent
of change: what’s the mechanism? 50
10 Weber, Grice, Crocker: Effects of prosody on the resolution of word-order ambiguities 51
11 Bard, Anderson, Flecha-Garcia, Kenicer, Mullin, Micholson, Smallwood, Chen:
Con-trolling attention and structure in dialogue: the interlocutor v the clock 52
12 Carlson, Kennedy, Dickey: Accents, structure, and the interpretation of gapping sentences 53
13 Catchpole, Hartsuiker, Pickering: Self-corrections in speech: Evidence against Levelt’s
Main Interruption Rule 54
14 Eiter, Radach, Inhoff: The Nature of the Phonological Code Accessed Early in Visual
Word Recognition during Sentence Reading 55
15 Longtin, Hall´e, Segui: Word processing, morphological surface structure and the lexicon 57
16 Tsai, Chen, Chen: Implicit Priming with Direct Word Naming in Single Word Production 58
17 Yang, Gordon: The Computational Cost of Syntactic and Semantic Integration for
Pro-cessing Sentences with Relative Clauses — A Case Study of Chinese 59
18 Hsu, Bruening: Investigating Gap-Filler Dependencies in Chinese: Is There an ‘Active
Gap’? 61
19 Cacciari, Padovani, Verzellesi, Carreiras: Contextual effects on the time course of
lin-guistic and conceptual gender information in understanding double-gender words 62
20 Bradley, Fern´andez, Lovric: Overt prosody in the RC-attachment construction:
Elicita-tion protocols 63
21 de Goede, Wester, Bastiaanse, Swinney, Shapiro: Verb activation patterns in Dutch matrix
clauses during on-line spoken sentence processing 65
22 Janssen, Muckel, Schiesser: Frequency can account for word order effects in German:
A matter of the appropriate corpus query 66
23 Hai: Computer Simulation of a Serial Parsing Model 67
24 Roland, Elman, Ferreira: Why ‘that’? 69
25 Kruijff, Vasishth: Quantifying word order freedom in natural language: Implications for
sentence comprehension 70
26 Mariol, Schelstraete: Implicit learning and lexical spelling: Rules abstraction and/or
frequency based processes? 71
27 Fossard, Garnham, Cowles: Referential Accessibility and Anaphoric Resolution: the
Case of the Demonstrative Noun-Phrase That N 72
28 Corley: Filled pauses can um help the listener 73
29 East: Simulating spell-out-of-trace with an SRN 74
Trang 430 Michael, Gordon: Differential processing of sentential information: Effects on Recovery
from the Garden Path 75
31 Roberts, Chater: Spreading activation as a mechanism for lexical smoothing 76
32 Santesteban, Costa: Can highly-proficient early bilinguals acquire two independent
syn-tactic systems?Influence of L1 Head-Parameter on DP production in L2 77
33 Poesio: Combining salience and lexical filtering for bridging references resolution 78
34 Tanaka: Dual-mechanism model in derivational morphology - evidence from Japanese
Causative formation 79
Poster Session 2, Tuesday August 26th
1 East: The parser: a word-level thief? 82
2 Hemforth, Schimke: Length and position in French 83
3 Hemforth, Konieczny, Bueche: Who was in France? The accessability of referents in
RC-attachment 84
4 Konieczny, D ¨oring: Anticipation of clause final heads Evidence from eye-tracking and
SRN simulations 85
5 Pynte, Colonna, Gola: Can an “unnatural” prosody help parsing? 86
6 Branigan, Cleland, McLean: Modality effects on sentence production: syntactic activation
in speech, writing and typing 87
7 Nicholson, Bard: The Intentionality of Disfluency : Findings from Feedback and Timing 88
8 Carlson, Dickey, Frazier, Clifton: Sluicing is affected by focus, but how? 89
9 Christianson: When Reanalysis Becomes Too Costly: Heuristic-based Parsing in a Free
Word-Order Language 90
10 Agnihotri, Mahajan, Vasishth, Fanselow, Schlesewsky: More isn’t always better: A
sur-prising constraint on retrieval cues in human sentence parsing 91
11 Gaskell, Cox, Foley, Grieve, O’Brien: To “thee” or not to “thee”? Constraints on
allo-morph pronunciation 93
12 Cardillo: Semantic Priming by a Sentence Context in Bilinguals: Reduced Context
Sen-sitivity and Inhibition 94
13 Torgersen: Temporal and semantic factors in perception of the voicing contrast in L2:
Effects of L2 proficiency on processing strategies 95
14 Andonova, Janyan: Searching for Gender Congruency in Bulgarian 96
15 Kaiser, Trueswell: Beyond the basic pronoun: A look at the referential properties of
Dutch anaphoric paradigms 97
16 Heuttig, Altmann: Language-mediated eye movements and the differential effects of the
processing of perceptual and conceptual colour information 99
17 van Gompel, Pearson: Competition during lexical ambiguity resolution 100
18 Kreiner, Koriat: From Structure to Meaning: The Contribution of Prosodic representation 101
19 Grodner, Sedivy: The Effect of Speaker-Specific Information in On-Line Pragmatic
In-ferencing 102
20 Yamashita, Chang, Hirose: Language-dependent Aspects of Structural Priming 103
21 H¨artl: No repetition priming with conceptually plausible objects: Object categorization
in sentential contexts 105
22 Cowles, Garnham: Antecedent Focus and Conceptual Distance in NP Anaphor Resolution106
23 Steiner: Recycling Structure: A New Dimension in the Processing of Coordination 107
24 Thomson, Shillcock, McDonald: The Role of the Magnocellular Pathway in Visual Word
Recognition 109
25 H¨aussler, Bader, Bayer: Number attraction in sentence comprehension 110
26 Ashby, Martin, Morris: Syllable effects in English word recognition: An ERP investigation112
27 Mirkovic, Seidenberg, Joanisse: Leaving the past tense behind: computational studies of
a morphologically-complex language 113
28 Boland, Hartsuiker, Pickering, Postma: The implementation of repairs in speech
pro-duction: evidence from the time-course of different repairs 114
29 Seegmiller, Townsend, Ingraffea: The Role of Event Structure in Comprehending Garden
Path Sentences 115
Trang 530 Betancort, Gillon Dowens, Carreiras: Subject and Object relative clauses: Whats going
on in Spanish? 116
31 Leicht: Why some people do not experience the garden path effect? The case of Hebrew 117
32 Monaghan, Chater, Christiansen: Phonological typicality and lexical processing 118
33 Buttery, Villavicencio: Language Acquisition and the Universal Grammar 119
34 Ward, Sturt, Sanford, Dawydiak: The effects of linguistic focus and semantic distance
on eye movements in a text-change detection task 120
Poster Session 3, Wednesday August 27th
1 Hemforth, Konieczny, Schimke: Modifier attraction and object attraction 122
2 Pynte, Gola: Adjunct predicates, discourse context, and the primary/non-primary distinction124
3 Christianson: Interpreting ‘pro’ in Isolated Sentences 125
4 Pickering, Branigan, McLean: Dialogue structure and the activation of syntactic information126
5 Vasishth: The role of decay and activation in human sentence processing 127
6 Shillcock, Monaghan: Visual word recognition and the anatomy of the visual system: the
processing style of the corpus callosum 129
7 Pitel, Sansonnet: Toward a Uniform Architecture for Processing Application-Oriented
Dialogue 130
8 Featherston: The relationship between judgement data and frequency data in syntactic
well-formedness: The Decathlon Model 132
9 Muckel, Schliesser, Pechmann: Prosodic cues to argument structure in German 133
10 McDonald, Thomson: Using eyetracking to contrast lexical-level and conceptual-level
connections in the bilingual lexicon 134
11 Vainio, Hy¨on¨a, Anneli Pajunen: Relative co-occurrence frequency affects the processing
of verb arguments in Finnish: An eye-tracking study 135
12 Gennari, MacDonald: Linking production and comprehension in processing relative
clauses 136
13 Repp, Sommer: Verb coding in complex sentences: Evidence from eye-movements in
the production of coordination with and without ellipsis 137
14 Warren, Gibson, Jameson, Hirsch: Effects of NP type on reading English clefts 138
15 Desmet, Ferreira: Implicit Causality as an Inherent Feature of Verbs and Verb Classes 139
16 MacLeod, Garrod: Where is the Twist in the Tongue Twister? Dysfluencies following
Spoken and Heard Twister Strings 140
17 Chen, Chen: Morphological Processing in the Production of Chinese Compound Words 142
18 Sakas, Fodor: Slightly ambiguous triggers for syntactic parameter setting 143
19 Shen, Mitchell: A Poster of Relative Clause Attachment Processing in Chinese 144
20 Trueswell, Kaiser: Information Source and Temporal Precedence in Syntactic Ambiguity
Resolution: Evidence from Verb-Final Constructions in Dutch 145
21 Igoa, Teira: The contribution of prosody to relative clause attachment in spoken
sen-tences in Spanish 147
22 Kiakogiorgi: Evaluating the morphological competence of school - age children in a
highly inflected language: a crosslinguistic perspective 148
23 Wardlow, Ferreira: Finding Common Ground: How Do Speakers Block Privileged
In-formation? 150
24 Irmen: How do grammatical and stereotypical gender influence the representation of
person information? 151
25 K¨uhnast, Saddy, Stojanova: Processing Negation and Aspect in Bulgarian Evidence from
Normal and Agrammatic Sentence Comprehension 152
26 Kempen, Harbusch: A corpus study into word order variation in German subordinate
clauses: Animacy affects linearization independently of grammatical function assignment 153
Trang 6Enriched compositional processing
Brian McElree1, Martin J Pickering2, Matthew J Traxler3, Steven Frisson4
brian.mcelree@nyu.edu, Martin.Pickering@ed.ac.uk, mjtraxler@ucdavis.edu
1 New York University
2 University of Edinburgh
3 University of California at Davis
4 University of Massachussetts at Amherst
Although traditional views of composition hold that semantic properties retrieved from the lexicon
are simply combined in a manner informed by syntactic structure, recent formal analyses and online
stud-ies of common expressions suggest that a more complex mechanism is needed to compute contextually
appropriate interpretations Compositional processes can modify default interpretations of expressions
and introduce semantic content not explicitly represented in the sentence or discourse One class of
expressions requiring an enriched form of composition is VPs involving verbs that select for event
com-plements (“begin”, “enjoy”, etc.) followed by NP comcom-plements that denote entities (“the book”, “a beer”,
etc.) Self-paced and eye-tracking measures have demonstrated that these expressions are more difficult
to process than control expressions, which can be interpreted with simple compositional operations that
combine lexically stored senses (McElree et al., 2001; Traxler et al, 2002)
We report three new eye-tracking studies that examined how contextual information affected
en-riched composition in order to identify what specific operations distinguish enen-riched from simple
com-position An expression like “the student began the book”, which readers report interpreting as “the
student began reading the book”, may be difficult to process because it requires complex operations to
construct a specific event sense for the complement, it is ambiguous and engenders competition between
alternative interpretations (e.g., reading versus writing the book), or it requires a costly retrieval
oper-ation to recover a plausible activity for the event interpretoper-ation Introducing the activity with a context
like “The student was reading in the library” did not attenuate the processing cost: With this context or
with a neutral context (“The student was resting in the library”), a target sentence like “After a while, he
started a book ” was more difficult to process (e.g., longer total times on the verb and complement NP)
than a control sentence like “After a while, he read a book ” However, introducing the entire event
sense with a context like “The student started a book in his dorm room” or “The student read a book in
his dorm room” completely eliminated the difficulty, whether the target sentence repeated the NP, “ he
started a book ” (Experiment 2) or whether it replaced the NP with a pronoun, “ he started it ”
(Experiment 3)
These findings are incompatible with ambiguity- or retrieval-based accounts and suggest that
inter-pretation is costly when composition requires construction of a sense not lexically stored or available in
the immediate discourse
References
McElree, B., Traxler, M J., Pickering, M J., Seely, R E., & Jackendoff, R (2001) Reading time evidence
for enriched semantic composition Cognition, 78, B15–B25
Traxler, M., McElree, B., & Pickering, M (2002 ) Coercion in sentence processing: Evidence from
eye-movements and self-paced reading Journal of Memory and Language, 47, 530–547
The representation of transitivity information
Manabu Arai1Roger P.G van Gompel1Jamie Pearson2Vera Schumacher3
m.arai@dundee.ac.uk, r.p.g.vangompel@dundee.ac.uk, jamie.pearson@ed.ac.uk
1 a The youngster and the schoolgirl were bullying (intransitive, verb repeated)
b The youngster and the schoolgirl were jeering (intransitive, verb not repeated)
c The youngster was bullying the schoolgirl (transitive, verb repeated)
d The youngster was jeering the schoolgirl (transitive, verb not repeated)
2 While the prisoner was bullying
According to the lexically specific hypothesis, priming should be larger when the verbs in the prime andtarget are the same than when they are different, because transitivity information is represented separatelyfor different verbs The category-general hypothesis predicts that priming should be unaffected by verbrepetition, because transitivity information is only associated with the class of verbs as a whole
We observed an interaction between transitivity of the prime and verb repetition After intransitiveprimes, participants produced more intransitives when the target verb was repeated than when it was notrepeated This indicates that information about intransitives is represented at the lexically specific level
In contrast, no verb repetition effect occurred after transitive primes, indicating that information abouttransitives is represented at a category-general level
In Experiment 2, we investigated how people represent transitivity information in passive sentences
If transitivity information is represented separately for passives and actives, transitive priming from sives (3a) should be smaller than from transitive actives (3b) In contrast, if transitivity information forpassives and actives is represented together, transitive priming from passives should be the same as fromactives
pas-3 a It is unimaginable to the youngster that the schoolgirl was bullied (passive)
b It is unimaginable that the youngster has bullied the schoolgirl (active)Participants produced fewer transitive sentences after (3a) than after (3b), indicating that transitivityinformation for actives and passives is represented separately Furthermore, in line with Experiment 1,
we observed no effect of verb repetition
We modify Pickering and Branigan’s (1998) syntactic representation model to explain the resultsfrom both experiments Most importantly, we claim that information about transitives is represented dif-ferently from information about intransitives We argue that the reason for this difference is that thetransitive subcategorisation frame represents the default case, because almost all verbs can be used tran-sitively (even so-called intransitive verbs like sneeze: Peter sneezed blood) As a result, there is no need
to represent information about transitives for individual verbs In contrast, not all verbs can be usedintransitively, and therefore information about intransitives is represented for individual verbs
Trang 7Processing Idioms in the Left and Right Hemispheres
Laurie A Stowe, Monika, Z Zempleni and John Hoeksl.a.stowe@let.rug.nl, m.z.zempleni@let.rug.nl, j.c.j.hoeks@let.rug.nl
Rijksuniversiteit Groenigen
Idioms are fixed phrases which appear to be stored in the mental lexical with their meaning, which
is usually not compositionally derived from the individual words Right hemisphere brain damage
fre-quently affects idiom comprehension Tompkins, Boada and McGarry ((1992) demonstrated that RH
patients showed speeded word monitoring within idioms as opposed to non-idioms even if they were
unable to paraphrase idioms, suggesting that idioms can be accessed as sequences by these patients but
that the meaning cannot be used normally The exact contribution of the right hemisphere in dealing with
the meaning of these sentences remains unclear
In the current study, we investigated the role of the two hemispheres by presenting 54 Dutch
sen-tences which were disambiguated by preceding context to the idiomatic or the non-idiomatic meaning
(e.g The minister stood on her toes, meaning annoyed her, vs The ballerina stood on her toes, which has
the compositional interpretation) These sentences were presented one word at a time in the center of a
computer screen, in order to be available to both hemispheres At the end of each sentence, subjects were
presented with an unrelated word, a word which was related to the sentence’s idiomatic meaning (e.g
annoy) or a word which was related to the literal meaning (e.g dance) in either the left or right visual
field, where it could initially only be processed by the the contralateral hemisphere (divided visual field
technique) Subjects made a lexical decision to the target words
After idiomatic sentences, there was a significant interaction between relatedness and visual field;
for words presented to the left hemisphere there was priming of the idiom related word relative to both
unrelated words and literal related words; the latter did not differ Words related to either meaning showed
facilitation in the right hemisphere This evidence suggests that the left hemisphere tends to choose the
idiomatic meaning, while the right hemisphere keeps both options open, as it does for other types of
ambiguity (Faust and Gernsbacher, 1993)
After literal context sentences, on the other hand, both idiom and literal related words were slowed
relative to the unrelated word in the right hemifield (left hemisphere), suggesting that needing to inhibit
the idiomatic meaning caused considerable processing difficulties On the other hand, the right
hemi-sphere showed faster decision times for literal related words than for unrelated words These results
clearly confirm that the two hemispheres respond to the disambiguation of these sentences in a
qualita-tively different way
References
Faust ME, Gernsbacher MA Cerebral mechanisms for suppression of inappropriate information during
sentence comprehension Brain Lang 1996; 53: 234-259
Tompkins,C.A, Boada, R., and McGarry, K 1992 The access and processing of familiar idioms by
brain-damaged and normally aging adults Journal of Speech and Hearing Research 35(3): 626-637
Grammatical gender and mapping of referential expressions in Russian
Irina Sekerinasekerina@postbox.csi.cuny.edu CUNY College of Staten Island
For many languages, it has been shown that in production, gender-matching prenominal adjectivescan facilitate noun identification (Bates et al., 1996) In comprehension, the research on the role ofgender has produced mixed results (JPR, 1999) In an eye-tracking experiment with gender manipulation
in French, Dahan et al (2000) did not find gender priming (le zebre vs le balai/la chaussette) althoughthe cohort effect was eliminated by the gender-marked article (le bouton vs la bouteille) Dahan andcolleagues proposed two competing explanations for the role of gender: grammar-based (grammaticalgender carried by the article) vs form-based (article co-occurrence with the noun) Based on their resultswith the article-noun and adjective-noun pairs in French, they argued for the form-based effects of gender
in spoken-word recognition
We present the results from a head-mounted eye-tracking experiment in Russian (N=16) thatclearly supports the immediate influence of grammatical gender encoded on the color adjective inconstraining the referential set of nouns In Russian, a noun’s gender (MASC, FEM, or NEUT) mor-phologically marked by the noun ending is also encoded in modifying adjectives - “krasn-aja / yj /oe” (’red+FEM /MASC /NEUT’) Three types of visual displays were manipulated: Control (the redcar+FEM), with Gender-mismatching distractor (the red car+FEM and the red flower+MASC), and withGender-matching competitor (the red car+FEM and the red squirrel+FEM.) The second factor manipu-lated the adjacency of the adjective and the noun: canonical ADJ NOUN VERB vs discontinuous ADJVERB NOUN (’The red put car in Position 6’), common in Colloquial Russian
In contrast to the Dahan et al.’s results, we found that in the canonical ADJ+NOUN word order,there were significantly more looks to the gender-matching competitor (the red squirrel) than to gender-mismatching distractor (the red flower), 35% vs 5% These looks occurred in the noun region, 300-700
ms after the onset of the noun but before its offset Moreover, the effect of gender was significant even inthe discontinuous ADJ NOUN order, when only the gender-matching competitor was rapidly checkedand discarded resulting in target fixation well before the noun was encountered, during the verb region.Looks at Competitor Gender-mismatching Gender-matching
Trang 8Situating language in the visual world: Mapping the semantics of
language onto the semantics of visual scenes
Gerry T.M Altmann1, Silvia P Gennari2and Louise A Shackleton3
g.altmann@psych.york.ac.uk, sgen@lcnl.wisc.edu, las111@york.ac.uk
1 University of York
2 University of Wisconsin-Maddison
Static scenes can be interpreted and encoded in respect of how the portrayed state may change
(cf Freyd, 1987) In principle, a visual scene portraying a man, a young girl, a motorbike, a carousel,
some beer and some sweets, may engender representations that encode the possibility that riding (of the
motorbike or carousel) may take place Subsequently, if participants hear the man/girl will ride/taste
(cf Kamide et al., 2003), aspects of the semantic event structures that would result from interpreting this
sentence may already have been constructed during the interpretation of the scene We report here two
studies which assessed whether visual scenes can prime visual (Expt 1) or spoken (Expt 2) words, and
specifically, verbs
In Experiment 1, 50 participants were shown 32 scenes showing, for example, a girl dancing (the
related condition), or sitting on the floor (the unrelated condition) The scene stayed onscreen for 3000
ms and was followed by a fixation prompt and then a visual word (e.g DANCE) in response to which
participants had to make a lexical decision Mean decision latencies were 571 ms and 547 ms in the
unrelated and related conditions respectively (p<.0001) We conclude that portrayed actions facilitate
recognition of words that refer to those actions
In Experiment 2, we took the stimuli from an existing study (Experiment 2, Kamide et al., 2003) in
which participants heard the man/girl will ride/taste in the context of the fairground scene described
above The 48 verbs spliced from each sentence were used as the stimuli in an auditory lexical decision
task each was presented with either the corresponding scene from that study (related) or with a different
scene (unrelated) The scenes stayed onscreen for 1750 ms before presentation of the auditory stimulus
for lexical decision Mean latencies were 855 ms and 817 ms in the unrelated and related conditions
respectively (p<.003) The magnitude of the difference correlated on an item-by-item basis with the
eye movement record from the original study (with both the probability and timing of eye movements
towards the appropriate objects) Thus, the more the scene facilitated verb recognition (in effect, the
more constraining the context), the greater the likelihood of launching an eye movement towards the
appropriate objects
We conclude that visual scene interpretation activates semantic structures denoted by verbs, such
as participants and actions, and that scene interpretation and language interpretation thus share semantic
structure
References
Freyd, J L (1987) Dynamic mental representations Psychological Review, 94(4), 427-438
Kamide, Y., Altmann, G T M., & Haywood, S L (2003) The time-course of prediction in incremental
sentence processing: Evidence from anticipatory eye movements Journal of Memory and Language
Dividing up referential labor: Finnish pronouns and demonstratives in
In Finnish, human referents are referred to with the gender-neutral pronoun ‘h¨an’ (he/she) or thedemonstrative ‘t¨am¨a’ (this, he/she) (e.g.[4]) To investigate the referential properties of these expressions,participants‘ (N=16) eye-movements were recorded while they listened to short stories and looked at pic-tures about the stories The participants’ task was to correct any mistakes in the stories Each target itemcontained a subject-verb-object or object-verb-subject sentence, followed by a sentence beginning with
‘h¨an’ (he/she) or ‘t¨am¨a’ (this, he/she) (ex.1) All critical pictures contained two human referents Therewere four conditions: (1)[SVO.H¨an], (2)[SVO.T¨am¨a], (3)[OVS.H¨an], (4)[OVS.T¨am¨a] The description
in the critical sentence was incorrect with respect to both pictured referents, in order to prompt ipants to correct the sentence, thereby providing another measure of how they resolved the anaphoricforms
partic-If we assume that pronouns and demonstratives map onto an accessibility scale, we expect thepronoun ‘h¨an’ to refer to the most accessible referent, and the demonstrative ‘t¨am¨a’ to the less accessibleone However, the eye-movement patterns show that these forms are not mirror images of one another
‘H¨an’ prompts significantly more looks to the subject–in both SVO and OVS conditions–than ‘t¨am¨a.’However, it is not the case that ‘t¨am¨a’ triggers looks to the object in SVO and OVS conditions Instead, for
‘t¨am¨a’, eye-movements show that listeners eventually opt for the second-mentioned character, regardless
of grammatical role
In conjunction with off-line sentence-completion data that we collected, these findings suggest that:(1) the demonstrative ‘t¨am¨a’ tends to refer to a second-mentioned character, and (2) the pronoun ‘h¨an’tends to refer to subjects, regardless of word order We hypothesize that these two expressions differ
in terms of the level of representation that they access: ‘h¨an’ taps into the syntactic level and is usedfor subjects, and ‘t¨am¨a’ looks at the discourse/pragmatic level and is used for low-salience referents
In sum, the referential properties of ‘h¨an’ and ‘t¨am¨a’ are not subject to a single common factor, and
an accessibility scale is not sufficient to explain the division of ‘referential labor’ between these twoexpressions
Example(1): Liisa is in the park She sees a man near a lake Then
man-subject greets woman-object ‘H¨an’ is wearing a red hat (SVO.H¨an)man-subject greets woman-object ‘T¨am¨a’ is wearing a red hat (SVO.T¨am¨a)man-object greets woman-subject ‘H¨an’ is wearing a red hat (OVS.H¨an)
Trang 9man-object greets woman-subject ‘T¨am¨a’ is wearing a red hat (OVS.T¨am¨a)
References
[1]Ariel, M (1990) Accessing NP Antecedents London: Routledge/Croom Helm
[2]Gernsbacher, M.A.(1990) Language Comprehension as Structure Building LEA
[3]Gundel, J.K., Hedberg, N & Zacharski, R (1993) Cognitive Status and the Form of Referring
Expres-sions in Discourse Language 69:274-307
[4]Halmari, H (1994) On Accessibility and Coreference NJL 17:35-59
[5]Hoffman, B (1998) Word Order, Information Structure and Centering in Turkish In Centering Theory
in Discourse Oxford: Clarendon Press 251-272
The role of competition in the production of referring expressions
Jennifer Arnold1and Zenzi Griffin2
jarnold@bcs.rochester.edu, zg9@prism.gatech.edu
1 University of Rochester
2 Georgia Institute of Technology
A central part of speaking involves deciding how to refer to people and things, for example choosingbetween a name and pronoun Most explanations of this choice suggest that pronouns are used for entitiesthat are salient in the discourse (e.g., Grosz et al, 1995; Gundel et al., 1993; Levelt, 1989), proposing asingle, uniform category of most salient, or in focus Centering Theory goes so far as to suggest that only
a single entity can occupy this privileged position In addition, pronouns are used more when there isonly one entity matching pronominal features like gender (Karmiloff-Smith, 1985) This Gender Effect
is most commonly explained in terms of a desire to be unambiguous (e.g., Chafe, 1994) We presentresults from a production study that challenge the notion of a single most-salient category, suggestingthat the salience of even the top-ranked entity is vulnerable to competition from the presence of additionalanimate characters, even when they have a different gender
Subjects viewed panel 1 of a 2-panel cartoon while they heard and repeated the first line of a storyabout it, e.g Mickey went for a walk with Donald/Daisy in the park one day They then generated
a second line in response to panel 2 In Exps 1a-c, both panels presented two characters, of eithersame or different gender We also analyzed filler items with only a single character in both panels, e.g.Minnie went snorkeling during her vacation in the Bahamas We only examined references to the subjectcharacter, the top-ranked and most likely character to be referred to with a pronoun Results showed morepronoun use in the different-gender (45%) than same-gender context (23%), but even more pronouns inthe one-character filllers (84%)
Exp 2 explicitly contrasted one- and two-character contexts, finding higher pronoun use when therewas only one character (63%) compared to two (23%), even though the two had different genders Onlythe number of characters in the first sentence mattered, and not whether panel 2 had one or two Thissuggests that what is important is the salience of these characters in the discourse model, and not thedecision about who to mention
This study provides novel evidence that the presence of another character impedes pronoun use,even when ambiguity avoidance is not an issue Together with the gender effect , this suggests thatsalience is modulated by competition among similar entities in the discourse model
References
Chafe, W 1994 Discourse, consciousness, and time Chicago: Chicago University Press
Grosz, B J., Joshi, A K & Weinstein, S 1995 Centering: A framework for modelling the Local Discourse.Computational Linguistics, 212
Gundel, J K., Hedberg, N & Zacharaski, R 1993 Cognitive status and the form of referring expressions.Language, 692.274-307
Karmiloff-Smith, A (1985) Language and cognitive processes form a developmental perspective guage and Cognitive Processes, 1(1), 61-85
Lan-Levelt, W J M 1989 Speaking Cambridge: MIT Press
Trang 10Referential Domains in Language Processing
Michael K Tanenhausmtan@bcs.rochester.eduDepartment of Brain and Cognitive SciencesUniversity of Rochester
Definite referring expressions, such as “the upcoming conference” assume a uniquely identifiable
referent within a relevant context or domain I’ll review recent work from my laboratory that uses
eye movements in relatively natural tasks to investigate how listeners circumscribe referential domains
as an utterance unfolds Factors widely viewed as relevant to defining referential domains within the
“language-as-action” tradition influence the earliest moments of reference resolution, including, salient
visual context, goal-specific properties of objects, and the listener’s knowledge of the speaker’s
per-spective, even when it conflicts with his or her own immediate perceptual experience I’ll highlight the
theoretical and methodological implications of this work, argue that further understanding of referential
domains requires investigating real-time processing of interactive conversation and review recent work
that shows how eye movements can be used to study moment-by-moment language comprehension in
fully interactive, non-scripted conversation
The interaction of mental representations from linguistic and visual input: A processing account of role-assignment in visual worlds
Pia Kn¨oferle1, Matthew W Crocker1, Christoph Scheepers2& Martin J Pickering3
{knoeferle,crocker}@coli.uni-sb.de, C.Scheepers@dundee.ed.ac.uk, M.Pickering@ed.ac.uk
1 Universit¨at des Saarlandes; 2 University of Dundee; 3 University of EdinburghWhile much work has examined referent selection and the effects of referential properties of ob-jects, little is known about the use of other ontological categories (e.g actions and events) and theirrole-assigning effects in sentence comprehension We report four studies using eye-movements in vi-sual scenes to investigate role-assignment and structural disambiguation through actions depicting role-relations between agents and patients in agent-action-patient events in initially structurally ambiguousspoken sentences
All experiments exploited similar pictures Images depicted 2 events each involving 2 of 3 ters One character was role-ambiguous (patient/agent), engaged in 2 events; the other two were unam-biguously agent and patient, respectively, e.g.:
charac-CELLIST ←splashing— BALLERINA ←sketching— FENCER
Unlike previous studies, visual stimuli offered only depicted role and event information tional/stereotypical knowledge was not available for disambiguation
Selec-Experiments 1 and 2 investigated the comprehension of initially structurally ambiguous spokenGerman SVO/OVS main clauses (NP1-V-ADV-NP2) While the NP1-V sequences alone did not disam-biguate, the depicted actions showed who-did-what-to-whom In Experiment 1, shortly after the verband before disambiguation by NP2 case marking, we observed anticipatory inspections to the patientfor SVO, and to the agent for OVS sentences In addition we found effects of initial role-interpretationand role-reinterpretation processes Experiment 2 reconfirmed the findings while controlling for intona-tion Experiment 3 established that depicted actions facilitated initial interpretation of role-relations inverb-final constructions
Experiment 4 replicated findings for German SVO/OVS clauses for a different construction(MV/RR) and language (English) We contrasted furthermore lexical (main verb/past-part) information(1) with purely functional (aux verb) information (2) in on-line disambiguation Sentences had the fol-lowing form:
1 a The ballerina splashed quickly the cellist in the white shirt
b The ballerina sketched quickly by the fencer splashes the cellist
2 a The ballerina will quickly splash the cellist
b The ballerina being quickly sketched by the fencer splashes the cellist
On the assumption that purely functional information may allow a more rapid disambiguation than lexical(verb) information, which only disambiguates together with event depictions, we predicted an earlierdisambiguation effect for sentences (2) vs (1)
Our findings suggest that depicted actions are used for rapid disambiguation of local structural androle ambiguity which verb knowledge alone could not have disambiguated We provide an account of theinteraction between linguistic and visual information in on-line sentence comprehension on the basis ofconceptual semantics (cf Jackendoff, 1983, 1990), bringing together insights from theoretical semanticsand experimental psycholinguistics
Trang 11Jackendoff, Ray 1983 Semantics and Cognition Cambridge, Mass.: MIT Press
Jackendoff, Ray 1990 Semantic Structures Cambridge, Mass.: MIT Press
Structure-Building and Lexical Access During Garden-Path Recovery
Fernanda Ferreira1, Janet Dean Fodor2, and Ellen Lau1
fernanda@eyelab.msu.edu, jfodor@gc.cuny.edu, lauellen@msu.edu
1 CUNY Graduate Center
2 Michigan State University
Many theories of garden-path recovery have been proposed, from back-tracking, to diagnosis, toretrieving an alternative analysis already computed However, there are few hard facts concerning theprocesses that occur between garden-path detection and garden-path repair We investigated this using
a forced rapid-judgment task (based on [1]) Participants read garden-path sentences (While Mary wasmending the sock fell off her lap), and similar decoy sentences which were not reanalyzable (Verb-decoy:While Mary was examining the sock fell off her lap; Prep-decoy: While Mary was looking at the sockfell off her lap) Presentation was visual, word-by-word, moving window At one of four time delays(D1=offset of disambiguating verb; D2=D1+325ms; D3=D1+650ms; D4=D1+975ms), a beep sounded,followed after 300 ms by a “response too slow” warning sound At the beep, subjects (N=40) made
a grammaticality judgment by pushing a button Afterwards they could either confirm or revise theirdecision at leisure
At delay D1, there were no significant differences in acceptance rates across sentence types tance of garden-paths then dropped for a while, through D3, but improved by D4 Most revealing was thecontrast in acceptance patterns for the two types of ungrammatical decoy In both, the subordinate-clauseobject was obligatory and therefore could not be stolen by the higher verb Acceptance of Prep-decoysdeclined rapidly to below 30% by D2 and continued on down For Verb-decoys, by contrast, accep-tance declined after D1 but leveled off from D2 to D3 before dropping at D4 For Verb-decoy sentenceswith higher-frequency subordinate verbs, there was even a temporary rise (“false optimism”) at D3 Thisparalleled the response pattern for the corresponding garden-path sentences; the two sentence types (un-grammatical vs grammatical) diverged sharply only at D4 These data suggest that the repair processstarts in all cases with an attempt to steal the ambiguous NP from the subordinate clause, followed bychecking the lexicon to verify that the revised structure is legitimate This lexical check is necessaryfor verbs, but not for prepositions if the parser assumes that prepositions as a class are transitive; soPrep-decoys are rejected more rapidly and categorically
Accep-These results are most consistent with the Diagnosis Model [2], which maintains that the parserengages in active reprocessing during garden-path reanalysis It does not simply abandon one analysisand switch to another, pre-formed structure Instead, it embarks on a course of action which includesstructural rearrangement, interleaved with lexical access as necessary
Trang 12Diagnosis and Garden-Path Recovery-Case and Agreement Symptoms
Revisited
Markus Bader & Josef Bayermarkus.bader@uni-konstanz.deUniversity of Konstanz
Why is recovery from a garden-path sometimes easy but sometimes difficult? Two major sources
of the processing difficulty caused by a garden- path effect are the costs of diagnosis and the costs of
cure or repair (cf Sturt & Crocker, 1998) According to the diagnosis hypothesis proposed by Fodor &
Inoue (1998), diagnosis is the most important contribution to garden-path strength The strongest test of
the diagnosis hypothesis would be a comparison where repair is held constant across conditions and only
the symptom signalling the need for reanalysis varies We have conducted several experiments making
such a comparison
A first set of experiments tested locally ambiguous object-subject (OS) sentences as below These
sentences differ only in the symptoms by which clause-final disambiguation is achieved: Due to
morpho-logical properties, (1) is disambiguated by a case-mismatch only, (2) by a number-agreement mismatch
only, and (3) by a case and number-agreement mismatch
1 Ich habe geh¨ort, daß Fritz eine Lehrerin half
I have heard that Fritz-DAT a teacher-NOM helped
“I heard that a teacher helped Fritz”
2 Ich habe geh¨ort, daß Fritz ein paar Lehrerinnen halfen
I have heard that Fritz-DAT a pair teachers-NOM helped
“I heard that a couple of teachers helped Fritz”
3 Ich habe geh¨ort, daß Fritz einige Lehrerinnen halfen
I have heard that Fritz-DAT some teachers-NOM helped
“I heard that some teachers helped Fritz”
Experimental results show that case-disambiguation leads to stronger garden- path effects than
ei-ther number or case-and-number disambiguation The latter two conditions did not differ from each oei-ther
This is in contrast to findings by Meng & Bader (2000) that in wh-questions number- disambiguation
leads to stronger garden-path effects than case- disambiguation
A second set of experiments investigated OS-sentences as in (4) and (5) The crucial difference
between these sentences and the sentences above is that the sentences in (4) and (5) contain verbs for
which the OS-order is the unmarked, base-generated order, whereas the OS-structure in (1)-(3) is derived
by moving the object in front of the subject
4 Mir ist erz¨ahlt worden, daß Maria ein Buch gegeben wurde
Me is told been that Maria-DAT a book-NOM given was
“I have been told that a book was given to Maria.”
5 Mir ist erz¨ahlt worden, daß Maria einige B ¨ucher gegeben wurden
Me is told been that Maria-DAT some books-NOM given were
“I have been told that some books were given to Maria.”
In (4), disambiguation is by case alone, in (5) it is by case and agreement (4) and (5) elicited a cleargarden-path effect, but type of disambiguation had no effect at all
We will show how our new results as well as the prior results of Meng & Bader (2000) can beexplained within a variant of the diagnosis model which makes diagnosis crucially dependent on theHSPM’s usual feature checking routines We will present a linking and feature-checking algorithm whichautomatically predicts the differential effectiveness of case and number symptoms in different syntacticconstructions
Trang 13Brain responses to differences in the processing of informational and
contrastive focus
H Wind Cowles1, Maria Polinsky2, Marta Kutas2, and Robert Kluender2
H.W.Cowles@sussex.ac.uk, polinsky@ling.ucsd.edu, kutas@cogsci.ucsd.edu, rkluender@ucsd.edu
1 University of Sussex
2 University of California at San Diego
The distinction between informational and contrastive focus as separate categories of information
structure is widely accepted in theoretical linguistics (Kiss,1999; and others) It has also been
demon-strated in the psycholinguistic literature that contrastive focus is processed in much the same way as
sentence and discourse topics (Arnold 1998, Cowles et al., 2000) These lines of research lead to the
pre-diction that informational and contrastive focus should show noticeable differences in on-line processing
We investigated this hypothesis in an event-related brain potential (ERP) study Identical, violation-free
target sentences were preceded by discourse contexts that either introduced three referents independently
in presentational sentences (informational focus) or set up an explicit contrast set of these referents
(contrastive focus) Target sentences thus differed only in information structure status, informational vs
contrastive focus, as shown below
Our results showed that subjects processed the target sentences differently depending on the
infor-mation structure imposed by prior discourse First, after the contrastive context, a slow anterior negative
potential was elicited by the first three words of the target sentence This effect is likely due to the
oc-currence of the discourse-linked which-phrase (Pesetsky, 1987) at the end of the contrastive discourse
context The which-phrase activated a contrast set of possible discourse referents in working memory
that must be maintained until the question is resolved at the grammatical subject of the target sentence
Intra-sentential syntactic dependencies are known to elicit similar effects of anterior negativity when
wh-fillers are activated in working memory (Kluender & Kutas, 1993; and many others), and anterior
negativity has also been elicited in discourse referential dependencies in an inter-sentential context (van
Berkum, Brown & Hagoort, 1999) Second, again following the contrastive context, a late positivity was
elicited at the subject head noun (butcher) The nature of late positivities in sentence-processing contexts
has been somewhat controversial, with some arguing that it is an index of syntactic processes This
can-not be the case in this study, since the late positivity was elicited by only one of a pair of syntactically
identical target sentences In this case, the late positivity is thus more likely to index integration (Coulson,
King & Kutas, 1998) or categorization (Kok, 2001) effects related to more general aspects of language
processing
In sum, differences in the information status (i.e., informational vs contrastive) of the same NP
referents in identical lexical and syntactic contexts result in on-line processing differences indexed by
brain responses
Example stimuli:
Informational Focus Context:
The kitchen of a posh restaurant was filled with people trying to get orders filled Near the
door was a butcher and another person A group of cooks was clustered around a stove,
including a chef and a specialist There was a pot of soup in the corner that was almost ready
to be served Did anyone taste the soup?
Contrastive Focus Context:
A butcher, a chef and a specialist were in the kitchen of a posh restaurant They had started
up the business together It was successful, but they were very busy All of them wantedeverything to be perfect, but only one had time to stop and check the soup Which one tastedthe soup?
Target Sentence: After a moment, the butcher tasted the soup
References
Arnold, J (1998) Reference form and discourse patterns Ph.D Dissertation, Stanford University.Coulson, S J King and M Kutas (1998a) Expect the unexpected: Event-related brain response to mor-phosyntactic violations Language and Cognitive Processes, 13, 21-58
Cowles, H.W., M Walenski, M Polinsky, and R Kluender (2000) Information Structure in Language cessing: A Study of Topic, Focus, and Pronoun Anaphora CUNY Conference on Human SentenceProcessing, La Jolla, CA
Pro-Kiss, K (1998) Identificational focus versus information focus Language, 74(2), 245-273
Kluender, R and M Kutas (1993) Bridging the gap: Evidence from ERPs on the processing of unboundeddependencies Journal of Cognitive Neuroscience, 5, 196-214
Kok, A (2001) On the utility of P3 amplitude as a measure of processing capacity Psychophysiology, 38,557-577
Pesetsky, D (1987) Wh-in-situ: Movement and unselective binding In Reuland & ter Meulen (eds) Therepresentation of indefiniteness, 98-129 MIT Press: Cambridge, MA
van Berkum, J.J.A., C.M Brown, and P Hagoort (1999) Early referential context effects in sentenceprocessing: Evidence from event-related brain potentials Journal of Memory and Language, 41,147-182
Trang 14How semantic analysis can overrule syntactic analysis: an ERP study
Marieke van Herten, Herman H J Kolk and Dorothee J Chwilla
M.vanHerten@nici.kun.nl, Kolk@nici.kun.nl, Chwilla@nici.kun.nl
University of Nijmegen
In three recent experiments, we have observed centroparietally distributed P600 effects instead of
N400 effects to semantic reversal anomalies (Kolk, Chwilla, van Herten, & Oor, 2003) These anomalies
were created by reversing subject and object, thereby yielding implausible scenarios (e.g.,
word-by-word translation from Dutch: ‘The fox that on the poacher hunted’ Paraphrase: ‘The fox that hunted the
poacher’) These sentences are not selectional restriction violations because both foxes and poachers can
hunt as well as be hunted P600 effects are usually associated with a grammaticality violation or with
reanalysis following syntactic ambiguity N400 effects typically follow semantically unexpected items
The occurrence of P600 instead of N400 effects to our semantic reversal anomalies in syntactically
correct and unambiguous sentences therefore was unexpected
To explain these results, we propose that in our sentences, participants have a strong bias for the
semantically most plausible interpretation (i.e., that the poacher hunts the fox) This bias must be so
strong that it initially overrules the syntactic analysis Hence no N400 effect is obtained A conflict arises
between the semantically plausible role interpretation provided by the semantic heuristic and the
implau-sible role interpretation, provided by the parser This conflict causes readers to reanalyze the preceding
input, which yields the P600 effect
In the current ERP study, we further tested our hypothesis by substituting implausibilities by true
selectional restriction violations (e.g., word-by-word translation: ‘The fox that on the poacher shot’
Para-phrase: ‘The fox that shot the poacher’) On all current accounts, these should elicit N400 effects, perhaps
followed by P600 effects Our bias hypothesis, however, predicts otherwise Because the semantic bias,
which is even stronger than in our previous experiments, overrules the syntactic analysis, participants
will initially not detect the semantic anomaly We therefore expect no N400 effects to occur Instead,
P600 effects should occur, reflecting that participants reanalyze the input
Twenty-six participants read sentences for comprehension The violations occurred in the middle
of the sentences The sentences were presented in RSVP mode with an SOA of 645 ms The EEG was
measured from 5 midline and 22 lateral sites The major results were as follows: (1) The selectional
restriction violations elicited no N400 effects but P600 effects (2) As in our previous studies the scalp
distribution of the P600 effects were similar to that found for syntactic anomalies These results provide
further support for our claim that semantic analysis can indeed overrule syntactic analysis
References
Kolk, H H J., Chwilla, D J., van Herten, M., & Oor, P J W (2003) Structure and limited capacity in
verbal working memory: A study with event-related potentials Brain and Language, 85, 1-36
Combinatory semantic processing can occur independently of syntactic
di-1 *The pizza had been delivering the VIOLATION
2 The man had been delivering the ACTIVE CONTROL
3 The pizza had been delivered by PASSIVE CONTROLStimuli like 1 contain an anomaly at the verb, which is potentially either syntactic or semantic in nature,depending on the interaction of syntactic and semantic processing The syntactic cues in 1 unambiguouslyindicate an active-voice analysis and therefore require a logical subject role for the initial noun phrase.This interpretation is semantically anomalous (compare to 2) By contrast, a logical object interpretation
is plausible However, this interpretation requires passive -ED inflection at the verb rather than -ING(compare to 3) Thus, the syntactic cues in the string are ill-formed to support this interpretation
We investigated the interaction of syntactic and semantic processing for stimuli like 1 Syntax-firstmodels of sentence processing argue that combinatory semantic processing is preceded and guided by
a stage of purely syntactic analysis (Frazier, 1987; Hahne & Friederici, 1999) This implies that at theverb in 1, the semantically difficult logical subject assignment will be pursued, due to support fromsyntactic cues Constraint-based models emphasize the influence of semantic constraints on languageprocessing (Trueswell, Tanenhaus & Garnsey, 1993) However, much of the theory and data concerningthis issue focus on situations of syntactic ambiguity, where syntactic constraints are conceivably not
at their strongest Here we investigate whether semantic constraints can drive interpretation in the face
of unambiguous opposition from syntactic cues The hypothesis is that well-formed strings like 1 will
be processed as ill-formed, due to the inconsistency between the verb’s inflection and the semanticallydriven interpretation
ERPs at the verb elicited a robust P600 effect for Violation stimuli relative to controls No difference
in N400 amplitude was observed The P600 effect indicates syntactic processing difficulty The absence
of N400 effects suggests that the implausible logical subject interpretation was not pursued It appearsthat semantic constraints can drive the assignment of thematic roles, even when multiple assignmentsare logically possible Under the conditions observed here, the semantic commitment is so rapid andentrenched that syntactic cues are rejected when they conflict with the semantically supported interpreta-tion In spite of strong syntactic support for an implausible thematic role assignment, there is no evidencethat the assignment is pursued Thus, the semantic process observed here appears to be highly efficientand to exhibit a substantial degree of independence from syntactic guidance
Trueswell, Tanenhaus, & Garnsey (1994) Semantic influences on parsing: Use of thematic role information
in syntactic ambiguity resolution Journal of Memory and Language 33 p 285-318
Trang 15Precision Timing in Speaking
Zenzi M Griffinzenzi.griffin@psych.gatech.edu; Georgia Institute of Technology
Speakers take a mean 900 ms to name familiar objects (Snodgrass & Yuditsky, 1996), but name
durations are under 500 ms Nonetheless, people can name two objects without pausing between names
If they spoke after only preparing the first object name, they would not have enough time to plan the
sec-ond name while articulating the first Neglecting to buffer first names would require speakers to prolong
articulation or pause while completing preparation of second names Speakers could avoid pausing by
preparing both names completely before speaking, but they do not Instead, they take into account the
time they needed to plan a second object name when deciding when to begin saying the first name
(Grif-fin, in press) For example, although ‘chef’ and ‘chandelier’ have similar object-naming latencies when
named alone, speakers began saying, ‘chandelier-tank’ earlier than ‘chef-tank’ when avoiding pauses
Eye movements, speech onsets, and durations suggest that speakers knew that chandelier would take
more time to articulate than chef and that chandelier therefore provided more time to prepare the second
object name (see also Meyer et al., 2003)
New experiments show that speakers are flexible in deciding whether to buffer the first of two object
names When not avoiding pauses, speakers spoke short and long first names equally early Consequently,
they often paused between names, particularly when first names were short In Block 2, the speakers were
encouraged to avoid pausing between names Half of the object pairs were new and half repeated For
new pairs, speakers spoke significantly earlier when first names were long rather than short They spoke
repeated pairs over 100 ms earlier than new pairs Thus, modulation of speech onset varied with the
availability of the object names as well as the first names length
These results have consequences for understanding how production processes are coordinated in
time It has been argued that verbs must be partially prepared before speech onset based on longer speech
latencies for subject+verb utterances than subject or verb alone (Kempen & Huijbers, 1983; Lindsley,
1978) However, speakers also begin speaking later when fluently naming two objects rather than one
Thus, the results used to argue for verb selection prior to speech may have nothing to do with verbs
per se Instead, pre-speech planning of a second word may follow from the time needed to prepare and
articulate words fluently
References
Griffin, Z M (in press) A reversed word length effect in coordinating the preparation and articulation of
words in speaking Psychonomic Bulletin and Review
Lindsley, J R (1975) Producing simple utterances: How far ahead do we plan? Cog Psychol, 7, 1-19
Meyer, A S., Roelofs, A., & Levelt, W J M (2003) Word length effects in object naming: The role of a
response criterion Journal of Memory and Language, 48, 131-147
Kempen, G., & Huijbers, P (1983) The lexicalization process in sentence production and naming: Indirect
election of words Cognition, 14, 185-209
Snodgrass, J G., & Yuditsky, T (1996) Naming times for the Snodgrass and Vanderwart pictures Behavior
Research Methods, Instruments, & Computers, 28, 516-536
Semantic facilitation and phonological interference in self-correction:
evidence from picture naming
Robert J Hartsuiker1, Martin J Pickering2, Nivja de Jong2
Robert.Hartsuiker@rug.ac.be, martin.pickering@ed.ac.uk, nivja.dejong@ed.ac.uk
1 University of Ghent; 2 University of Edinburgh
According to Dell, Burger, and Svec (1997), speakers deactivate lexical and phonological tations after having selected them An important function of that mechanism is to prevent speakers fromperseveration of material already said The present study asks whether deactivation also occurs followingthe detection of a speech error Do speakers deactivate the representation of the error (in order to preventthe error from reoccurring) or do they maintain a representation, in order to reuse it as much as possible(Hartsuiker & Kolk, 2001)?
represen-We tested these hypotheses in picture naming experiments where on a small proportion of trials thepicture was quickly replaced by another picture with a semantically or phonologically related name InExperiments 1 - 2, participants were instructed to consider the first picture as an ‘error’, to stop naming
it as soon as possible, and name the new picture instead (self-correction) Trials where participants terrupted and corrected after completely producing the name of the first picture (e.g., tiger lion) yieldedsemantic facilitation (Experiment 1) and phonological interference (Experiment 2) on the naming la-tency of the second picture But trials with word-internal interruptions of the first picture-name (tig lion)yielded semantic interference
in-These results suggest that after the actual production of an error, the lexical representations of thelemma and word form are deactivated but the conceptual representation is not This helps with semanticerrors, because there is less competition at the lemma level and there is facilitation from the conceptuallevel; but it hinders with phonological errors, because the correct phonological representations are rel-atively inactive (i.e., it is hard to say ‘sea shells’ after just having said ‘she sells’) But when the error
is only partially produced, the lemma of the error is not deactivated and competes with the lemma of asemantically related response
This interpretation was supported in Experiment 3 Here, the first pictures were surrounded byred or green frames and participants only named the second picture The speakers could predict thelikelihood that the picture changed from the frame’s color: if it was green (change-unlikely) 10% of thepictures changed, but if it was red (change-likely) 90% of pictures changed We analyzed the trials inwhich the picture indeed changed In the change-likely condition, the participants activated the concept(it is difficult not to recognize a picture) but presumably not the lemma This condition yielded semanticfacilitation In the change-unlikely condition, participants could safely assume that they were to name thatpicture Hence, we expected them to retrieve the lemma Consistent with that view, there was semanticinterference A fourth experiment, testing phonological relatedness, is currently underway
Trang 16How do speakers avoid linguistic ambiguity?
Victor S Ferreira , L Robert Slevc, and Erin S Rogers
ferreira@psy.ucsd.eduUniversity of California, San Diego
Expressions can be ambiguous, and ambiguity disrupts communication Thus, speakers should
avoid ambiguous expressions Some studies suggest they do (Snedeker & Trueswell, 2003), whereas
others suggest they don’t (Allbritton et al., 1996; Ferreira & Dell, 2000) Four experiments sought to
ac-count for this mixed picture by determining how speakers avoid ambiguities Speakers described displays
that included a target picture, a critical foil picture, and two filler pictures The target (e.g., a baseball
bat) was ambiguous with respect to the foil either linguistically (an animal bat), nonlinguistically (a
larger baseball bat), or not at all (a harmonica) Here, we focus on linguistic ambiguities Instructions
told speakers to describe targets so that they were distinguishable from the other pictures in their
dis-plays We measured how often target descriptions were ambiguous (“bat”) In Experiment 1, speakers
described targets either immediately upon display presentation, or after a five-second preview Speakers
were 24% less ambiguous with preview Experiment 2 explored whether this preview benefit happened
because prior to target description, speakers merely had time to see the pictures, or because speakers
had time to conceptually and/or linguistically encode the pictures A dot appeared and quickly bounced
from a filler to the target to the foil, and speakers had to describe either just the penultimately indicated
object (“[baseball] bat”) or all indicated objects (“windmill, [baseball] bat, [flying] bat”) Displays
dis-appeared when speakers began talking Speakers described pictures to real or hypothetical addressees
The penultimate-versus-all manipulation had no effect, and with real addressees, speakers were more
specific overall, but did not avoid ambiguity any differently Most importantly, even when describing
all objects, target descriptions were just as ambiguous as in Experiment 1 WITHOUT preview Thus,
seeing and/or preparing to describe objects cannot explain Experiment 1’s preview benefit Experiments
3 and 4 manipulated whether speakers described foils before or after targets When foils were described
after targets, speakers were as ambiguous as in Experiment 1 without preview, whereas when foils were
described before targets, they were as unambiguous as in Experiment 1 with preview Thus, speakers
described targets more unambiguously only when they had already described the ambiguity-causing foil
(Nonlinguistic ambiguities were nearly always avoided, and gave rise to none of these patterns.) The
results suggest that speakers avoid linguistic ambiguity by formulating linguistic expressions that they
then compare (before articulation) only to utterances already in the speech record Speakers’ ability to
compare currently formulated utterances to descriptions not yet in the speech record is extremely limited
References
Allbritton, D W., McKoon, G., & Ratcliff, R (1996) Reliability of prosodic cues for resolving syntactic
ambiguity Journal of Experimental Psychology: Learning, Memory, and Cognition, 22(3), 714-735
Ferreira, V S., & Dell, G S (2000) Effect of ambiguity and lexical availability on syntactic and lexical
production Cognitive Psychology, 40(4), 296-340
Snedeker, J., & Trueswell, J (2003) Using prosody to avoid ambiguity: Effects of speaker awareness and
referential context Journal of Memory & Language, 48(1), 103-130
Real-time circumscription of referential domains in non-scripted
of participants are presented here All reported effects are statistically significant
We replicated the standard cohort effect [2] when the cohort and target appeared on the same island.Approximately 56% of the time, participants specified which island the target was on before the onset
of the noun phrase In these constructions, the cohort effect was eliminated when the cohort and targetwere on different islands Cohort effects were also reduced when the speaker used an unmodified noun,e.g., “the cloud” but the competitors were members of a contrast set (e.g the island contained a largecloud [target], and a large clown & small clown [competitors]), demonstrating a clear effect of contrast
We expected a diminished cohort effect for references to targets in a contrast set (e.g “the large cloud”),when the cohort competitor (which was also large) did not have a contrasting member in the referentialdomain, i.e only one clown on that island [3] We did observe cohort effects in this condition, though,indicating that the presence of a scalar modifier did not completely restrict the domain of interpretation
to those entities with contrast members However, speakers frequently used scalar adjectives when thetarget was not a member of a contrast set and often used a comparative form when it was (e.g., “the largercloud”) Thus eye movement patterns correspond to the contingencies created by the participants.These findings demonstrate that (1) it is possible to track real-time referential domain circumscrip-tion in unrestricted conversation between na¨ıve participants, and (2) doing so sheds light on the mecha-nisms that support interactive conversation
[3] Sedivy, J.C., Tanenhaus, M.K., Chambers, C., & Carlson, G.N (1999) Achieving incremental semanticinterpretation through contextual representation Cognition, 71, 109-147
Trang 17What clarification requests tell us about dialogue information states.
Jonathan Ginzburgginzburg@dcs.kcl.ac.ukKing’s College London
Communication is frequently less than perfect The effect a speaker hopes to achieve doesn’t quite
go through And yet in the context of a dialogue, this does not result in failure, let alone a crashing
halt Rather, without further ado the (previous) addressee may pose a clarification request (CR), in which
she localizes her problem with the utterance, as illustrated in the following examples from the British
National Corpus:
1 a Unknown: He’s anal retentive, that’s what it is Kath: He’s what?
b Anon 1: you see the behind of Taz Selassie: Tazmania?
In this talk, I will discuss ongoing work I have been engaged in with some colleagues (including
Robin Cooper, Pat Healey, Matt Purver, and Ivan Sag) on a theoretical analysis of a number of classes of
CRs, as well as corpus-based and experimental investigation thereof CRs are interesting for a number of
reasons For a start, they provide crucial evidence for intrinsic asymmetries between conversationalists
in dialogue, that are potentially maintained in the ‘medium term’ at the very least Furthermore, CRs
provide strong evidence for the need to see contextual updates in terms that are not strictly semantic—
among other considerations is data concerning parallelism effects Given this, they raise obvious issues
about the persistence of structural information in context In addition, the forms used to express CRs
are frequently highly ambiguous, which raises various issues about the (not infrequently fallible)
disam-biguation strategies employed Last, but not least, the analysis of CRs requires grammar architectures
satisfying rather distinctive properties, not least the integration of highly complex semantic and
contex-tual information with structural information in a ‘fractal’ manner (i.e this applies uniformly as the parts
get smaller and smaller.)
Phonology impacts segmentation and generalisation in speech
processing
Luca Onnis1, Padraic Monaghan1, Nick Chater1, Korin Richmond2
L.Onnis@warwick.ac.uk
1 University of Warwick; 2 University of Edinburgh
Several theories of language acquisition contend that processing is dependent on either statistical oralgebraic computations Pe˜na et al (2002) proposed a reconcilist view: speech segmentation operates onthe basis of statistical learning, whereas entirely separate algebraic computations are necessary for learn-ing grammatical structure In a series of experiments we show that there is no evidence yet supportingthis segregation of computational processes
Pe˜na et al trained participants on continuous streams of syllables from an artificial language prised of words of the form Ai Xj Bi, with three such Ai Bi pairs, and Xj was one of three syllablesthat randomly intervened between the Ai Bi pair Participants preferred words (e.g., A1 X2 B1) overpart-words, i.e., sequences crossing word boundaries (e.g., X2 B1 A3 or B3 A1 X2) The nonadjacentdependencies between the Ai and the Bi syllables were learned and contributed towards segmentation.Pe˜na et al also tested participants on whether they learned to generalize from the rules of the stimuli.Participants demonstrated no preference for “rule-words”, composed of an Ai Bi pair with a different A
com-or B in the intervening position (e.g., A1 B3 B1), compared to part-wcom-ords
However, in a third manipulation where 25-ms gaps were introduced between words during thetraining phase, participants did generalize Pe˜na et al claimed that altering the speech signal resulted
in a change in the computations performed by their participants Statistical computations were used in
a segmentation task but this was not performed simultaneously with algebraic computations that wouldpermit generalizations of the structure
An alternative explanation to account for the results is that certain phonological properties of thestimuli were not controlled for thoroughly In each experiment, Pe˜na et al used syllables in the samepositions It is possible, then, that phonological properties exert an influence on the results rather thanthat participants learn the subtle statistical or algebraic properties of the stimuli We present below abattery of new artificial language experiments that manipulate the order and position of syllables, whichindicate that the confound of phonology is sufficient to account for all their results Consequently, there is
no evidence yet for learning, either statistical or algebraic, on the basis of the nonadjacent dependencies
in the stimuli
Because non-adjacent dependencies can be learned with either large variability of intervening items(G ´omez, 2002) or no variability (Onnis et al., submitted) we discuss that performance in artificial lan-guage experiments may depend heavily on the specific tasks and conditions constructed by the experi-menter
References
G´omez, R (2002) Variability and detection of invariant structure Psych Sci, 13, 431-436
Onnis, L., Christiansen, M.H., Chater, N., and Gomez, R.L (submitted) Active data selection for statisticalstructure in human sequential learning: Preliminary evidence from Artificial Grammar Learning.Pe˜na, M., Bonatti, L., Nespor, M., Mehler, J (2002) Signal-driven computations in speech processing.Science, 298, 604-607
Trang 18Multiple-Cue Integration in Language Acquisition: The Differential
Contribution of Phonological and Distributional Cues
Morten H Christiansen1; Florencia Reali1; Padraic Monaghan2; Nick Chater2
mhc27@cornell.edu, fr34@cornell.edu, padraic.monaghan@warwick.ac.uk, n.chater@warwick.ac.uk
1 Cornell University; 2 University of Warwick
In language acquisition, discovering syntactic constraints requires being able to assign individual
words to lexical categories, such as nouns and verbs However, lexical categories are only useful insofar
as they support syntactic constraints In this paper, we consider whether phonological and distributional
cues may be helpful for solving this ”bootstrapping” problem
Phonological information provides cues about the lexical category of individual words; e.g., in
English, nouns and verbs differ in terms of number of syllables, and position of stress (Kelly, 1992)
The distributional context of a word may also provide information about its potential lexical category;
e.g., mostly nouns occur after determiners (Maratsos & Chalkley, 1980) We quantified the potential
usefulness of phonological and distributional cues through a series of corpus analyses involving nouns
and verbs taken from 5,436,855 words of child-directed speech in CHILDES For the
phonological-cue analyses, we used 16 phonological phonological-cues suggested to be useful for distinguishing between different
lexical categories For our distributional analyses, we developed a novel, developmentally plausible
ap-proach in which we took the 20 most frequent words in the corpus, and counted the co-occurrence of
all other words following each of these context words The results show that the use of several cues
provides not only more accurate classification than single cues, but better generalization to novel
situa-tions We also found that frequency interacted with the two cues: distributional cues were more useful
for high-frequency words, whereas phonological cues were more reliable for low-frequency words The
integration of phonological and distributional cues resulted in 66.7% correct noun/verb classification
While the corpus analyses demonstrate that phonological and distributional information can provide
strong cues for differentiating between nouns and verbs, it is not clear whether a learning mechanism
would actually be able to utilize such probabilistic information We therefore conducted a set of simple
recurrent network (SRN, Elman, 1990) simulations in which networks were trained on the
Bernstein-Ratner corpus of child-directed speech Words were presented one-by-one, represented by the same 16
phonological cues as in our corpus analyses The results indicate that the networks were able to integrate
the distributional information learned from the corpus with the phonological cues, allowing the network
to go beyond the input with 81.1% correct noun/verb classification These results show that there are
learning mechanisms capable of integrating multiple probabilistic phonological and distributional cues
References
Elman, J.L (1990) Finding structure in time Cognitive Science, 14, 179-211
Kelly, M.H (1992) Using sound to solve syntactic problems: The role of phonology in grammatical
cate-gory assignments Psychological Review, 99, 349-364
Maratsos, M.P & Chalkley, M.A (1980) The internal language of children’s syntax: The ontogenesis and
representation of syntactic categories In K.E Nelson (Ed.), Children’s Language Volume 2,
pp.127-214 New York: Gardner Press
A Psychophysical Law for Linguistic Judgments
Frank Kellerkeller@inf.ed.ac.ukUniversity of Edinburgh
It has been argued that linguistic acceptability can be estimated using the psychophysical technique
of magnitude estimation (ME), in the same way as physical continua such as brightness and loudness(Bard et al 1996; Cowart 1997) For physical continua, plotting the perceived stimulus magnitude againstthe actual physical magnitude results in a power relationship, the Psychophysical Law (Stevens 1957).The crucial difference between ME of physical stimuli and ME of linguistic stimuli is that for thelatter, no objective standard of comparison is available: linguistic acceptability does not have a physicalmanifestation that can be measured directly The present research aims to address this problem, based
on the hypothesis is that the theoretical notion of number of constraint violations can form the basis of apower law for linguistic judgments analogous to Stevens’ Psychophysical Law
We present data from an ME experiment for a well-known syntactic phenomenon: word order ation in German This experiments elicits acceptability judgments for stimuli that contain between zeroand five violations of word order constraints We show how the word order data can be accounted for by
vari-a power lvari-aw vari-anvari-alogous to Stevens’ Lvari-aw thvari-at relvari-ates the number of constrvari-aints violvari-ated by vari-a sentence toits perceived acceptability This power law achieves a good fit on the experimental data (R = 89, N = 24,
p < 001) A linear law achieves a worse fit on the same data (R = 78, N = 24, p < 001) This difference
is significant, as shown by a t-test on the degree-of-freedom adjusted correlation coefficients
To show the generality of this finding, we apply our power law to seven different ME studies fromthe literature These published data sets cover a range of syntactic constructions (word order, extraction,gapping) in three different languages (German, Greek, English) Our results show that the power lawprovides a consistently good fit on the experimental data sets In all but two cases, the fit achieved by thepower law is significantly better than the fit of the linear law
Stevens (1957) lists 14 different modalities for which the Psychophysical Law holds; the exponent
of the power law is characteristic of a given modality, it can range from 3 for loudness to 2.0 for visualflash rate Our results show that the exponent for linguistic acceptability is around 36; this value seems
to be constant across data sets
Trang 19Anticipating Cause and Consequence during Text Comprehension
Julia Simner & Martin Pickeringj.simner@ed.ac.uk; martin.pickering@ed.ac.ukUniversity of Edinburgh
Generating expectations about the nature of upcoming text can aid comprehension when these
ex-pectations are found to be correct Such exex-pectations hold at multiple levels, such as for upcoming lexical
items (e.g., Ehrlich & Rayner, 1981) syntactic categories (e.g., Fodor, et al., 1974) and event-level
re-lations (Duffy, 1986; Simner, et al., 2003) This study investigates expectations at the level of event
structure, and asks whether readers expect upcoming discourse to give the cause, or the consequence
of previously described events Current studies are conflicting and have concluded, conversely, that the
expectation is both for causes (Majid, et al., 2003) and for consequences (Arnold, 2001; van den Broek,
et al., 2000) We examine the methodology and materials of these studies, and ask how readers’
expec-tations can be determined by features of the discourse In four experiments, we presented readers with
discourse contexts whose length and content were manipulated A discourse continuation task was used,
and participants’ continuations were coded for cause and consequence (both of necessity and sufficiency),
and these continuations used to determine readers’ expectations about upcoming text
We found that readers generate fine-grain causality expectations based on (a) text length (b) agent
typicality (c) task instructions, and (d) previous causal context More causal continuations were given
when the text was short (Experiment 1), and when it described atypical agents of events (e.g The
par-son/wrestler pushed Shaun; Experiment 2) We show also (Experiment 3) how previous task instructions
(van den Broek, et al., 2000) may have fallaciously inflated the proportion of consequential
continua-tions generated Finally, and of special interest, we found evidence of ’causal priming’ (Experiment 4)
in which previous cause or consequence information can prime readers to expect upcoming information
of a similar type We presented identical discourse fragments for continuation (e.g She applauded him)
that were immediately preceded either by a cause or by a consequence (e.g Beryl admired/delighted
John) The presentation of a cause led readers to anticipate further causes, while the presentation of a
consequence led them to expect further consequences Such evidence is surprising, and argues against a
(perhaps more intuitive) ’fill-in-the-gap’ approach, where readers seek information that is missing from
their mental model, (anticipating a consequence when a cause has previously been supplied, and vice
versa) To our knowledge, this is the first evidence of a priming mechanism in the domain of event-level
text processing, and it provides information about how readers make judgements about the likely content
of upcoming material
References
Arnold, J E (2001) The effect of thematic roles on pronoun use and frequency of reference continuation
Discourse Processes,31, 137-162
Duffy, S A (1986) Role of expectations in sentence integration.Journal of Experimental Psychology:
Learning, Memory, and Cognition, 12, 208-219
Ehrlich, S., & Rayner, K (1981) Contextual effects on word perception and eye-movements during
read-ing Journal of Verbal Learning and Verbal Behavior, 20, 641-655
Fodor, J., Bever, T., & Garrett, M (1974) The psychology of language New York: McGraw-Hill
Majid, A., Sanford A J., & Pickering, M J (2002) Do interpersonal verbs lead to focus on cause or
consequence? Unpublished manuscript
Simner, J., Garnham, A., & Pickering, M (2003) Discourse Cues to Ambiguity Resolution: Evidence from
’do it’ Comprehension.Discourse Processes (in press)
van den Broek, P., Linzie, B., Fletcher C., & Marsolek, C D (2000) The role of causal discourse structure
in narrative writing Memory & Cognition, 28, 711-721
Task demands and the allocation of attention in reading: some early
all-or-We report two experiments using materials like those in (1) If such sentences are fully interpreted,
we would expect the reading time for (1a) to be shorter than that for (1b), because the pronoun anaphorwithin the relative clause matches stereotypical expectations about the gender of the sentential subject(i.e., the engineer; cf Carreiras et al., 1996)
1 a The engineer who had lost his hard hat argued with the foreman
b The engineer who had lost her hard hat argued with the foreman
However, the pragmatic salience of the relative clause can be manipulated by preceding it with acontext introducing either one or two engineers Following Crain & Steedman (1985), we assume thatthe introduction of two engineers requires readers to interpret (1) as identifying a particular engineer(the relative clause is interpreted restrictively); if one engineer has been introduced, the modifier merelyadds extra information (nonrestrictive) Depending on task demands, (1a) and (1b) may exhibit differingprocessing costs in different contexts
In a self-paced reading task where no explicit manipulation of task demands was made, (1b) tooklonger than (1a) only following a two-engineer context (participants attended more to the relative clause
in order to establish reference) A second experiment manipulated the task demands by presenting perimental items in the guise of short newspaper stories, and asking participants to indicate their interest
ex-in the stories Eyetrackex-ing data from Experiment 2 revealed a difference only followex-ing a one-engex-ineer
context, where the pronoun in (1a) took longer to read than that in (1b) We offer a tentative explanation
of the detailed findings, together with the general conclusion that “what the reading is for” can greatly
impact the allocation of attention in language comprehension
Trang 20How the brain solves the binding problem for language
Peter Hagoortpeter.hagoort@fcdonders.kun.nlDonders Institute, University of Nijmegen
In my presentation I will discuss a series of ERP and imaging studies on sentence and discourse
processing The focus will be on both semantic and syntactic binding A neurocomputational model of
parsing will be proposed that accounts for both behavioral and ERP data on syntactic processing A
series of architectural principles of sentence and discourse processing will be discussed that are claimed
to follow from the empirical data Considerations of brain organization result in the proposal that the left
prefrontal cortex is a crucial area for binding syntactic and semantic information that is retrieved from
memory into a unified sentence/discourse level representation
Accessing antecedents: Pronouns with infrequent antecedents are easier
to process than pronouns with frequent antecedents
Roger P.G van Gompel1and Asifa Majid2
r.p.g.vangompel@dundee.ac.uk
1 Department of PsychologyUniversity of Dundee
2 Language & Cognition GroupMax Planck Institute for Psycholinguistics, Nijmegen
Although there is ample evidence that high-level linguistic factors such as discourse informationand sentence semantics influence the ease with which pronouns are processed (e.g., Garnham, 2002), it
is much less clear whether and how low-level, lexical factors influence the processing of pronouns Wereport an eye- movement reading study which showed that one low-level factor, the lexical frequency of
an antecedent, affects the processing of a subsequent pronoun
When readers process a pronoun, they need to reaccess at least some information about the tecedent in order to establish a coreference link There are three accounts of how antecedent frequencyinformation affects the processing of a pronoun On one account, reaccessing the antecedent involves thesame processes as initial lexical access According to this, pronouns with a high frequency antecedentshould be easier to process than pronouns with a low frequency antecedent Secondly, Simner & Smith(1999) argued that processing of pronouns does not involve reaccessing frequency information, and there-fore, there should be no effect of antecedent frequency at a pronoun Finally, a third account predicts thatfrequency has an effect on pronoun resolution through saliency When a noun is infrequent, readingtimes are long, and therefore an infrequent noun may be more salient and accessible than a frequentnoun (Pynte & Colonna, 2000) Given that pronouns are easier to process when they refer to a salient an-tecedent, this leads to the interesting prediction that pronouns that refer to infrequent antecedents should
an-be EASIER to process than pronouns that refer to frequent antecedents
In order to test these predictions, we used stimuli such as below Unambiguous pronouns (his) thatreferred to a high- frequent antecedent (actor) were contrasted with those that referred to a low-frequentantecedent (tenor)
The crowd thrilled the actor/tenor with a standing ovation They responded to his mance in an emotional way
perfor-Our results supported the saliency hypothesis First fixation and first-pass times for the antecedentwere longer when the antecedent was infrequent than when it was frequent, but this pattern was reversed
in both measures for the region following the pronoun (performance) We conclude that low-level, lexicalfactors influence the processing of pronouns: The less frequent the antecedent, the more salient it is, andthe easier it is to process a subsequent pronoun Furthermore, our data shows that reaccessing antecedents
is different from initial lexical access Finally, we will discuss why previous studies (e.g., Simner &Smith, 1999) failed to observe frequency effects at the pronoun
Trang 21Compound processing in Finnish: a survey
Raymond Bertram1, Jukka Hy¨on¨a1and Alexander Pollatsek2
rayber@utu.fi, hyona@utu.fi, pollatsek@psych.umass.edu
1 Dept of Psychology, Turku, FIN-20014 Turku, Finland;
2 Dept of Psychology, Massachusetts, Amherst, MA 01003, USA,
Over the past 5 years, we have conducted a number of studies on Finnish compound words to
acquire a deeper insight in several relevant aspects of lexical processing (Hy¨on¨a & Pollatsek, 1998;
Pollatsek, Hy¨on¨a, & Bertram, 2000; Bertram & Hy¨on¨a, 2003) In these studies we found that processing
long compounds involves both constituent and whole-word representations, whereas short compounds
are mainly recognized in a holistic manner On the basis of these findings, the following visual acuity
hypothesis was put forth: readers start to analyze the first constituent of long compounds simply because
they do not have enough letter information on the latter part of the word and 2) readers start to analyze
the whole-word string of short compounds immediately because all letters are within foveal vision The
current presentation focuses mainly on the findings of the more recent studies described below
The impact of constituents in processing long compounds implies successful segmentation of the
whole word Bertram, Hy¨on¨a, & Pollatsek (in preparation) investigated whether parsing long compounds
into their comprising constituents was aided by orthographic and/or phonological cues, making use of the
Finnish rules for “vowel harmony” In Finnish, vowels of different quality (front or back) cannot appear
in the same constituent Thus, SELK ¨A/ONGELMA ‘back problem’, with a first constituent ending in
a front vowel ( ¨A) and a second constituent starting with a back vowel (O), provides a clear cue as to
the location of the morpheme boundary, unlike SATU/OLENTO ‘fairytale creature’ with back vowels
throughout the word Processing of the compounds was aided when the boundary was made clear by the
vowel harmony constraints This segmentation cue was particularly helpful in resolving segmentation
problems, when the initial fixation was located relatively far away from the morpheme boundary In
contrast, when the initial fixation was around the morpheme boundary and thus in sharp focus, the parsing
mechanism worked smoothly even for compounds with only front or back vowels throughout the word
In the second study (Hy¨on¨a, Bertram, & Pollatsek, in preparation), we employed a variation of
the eye-movement contingent display paradigm in which a preview of the second constituent of the
compound was either present or partly replaced by orthographically similar letters before the second
constituent was fixated There was a large preview benefit when the second constituent was present, but
this effect mostly appeared only after the boundary was crossed and the second constituent was fixated
On the basis of these results, implications for models of eye movement behavior will be discussed
References
Hy¨on¨a, J., & Pollatsek, A (1998) Reading Finnish compound words: Eye fixations are affected by
com-ponent morphemes Journal of Experimental Psychology: Human Perception & Performance, 24,
1612-1627
Pollatsek, A Hy¨on¨a, J., & Bertram, R (2000) The role of morphological constituents in reading Finnish
compound words Journal of Experimental Psychology: Human Perception and Performance, 26,
820-833
Bertram, R., & Hy¨on¨a, J (2003) The length of a complex word modifies the role of morphological
struc-ture: Evidence from reading short and long Finnish compounds Journal of Memory and Language,
48, 615-634
Hy¨on¨a, J., Bertram, R., & Pollatsek, A Are long compound words identified serially via their constituents?
Manuscript in preparation
Bertram, R., Hy¨on¨a, J & Pollatsek, A Vowel disharmony as a segmentation cue in processing Finnish
compounds Manuscript in preparation
Lexical competition in visual word recognition: Evidence from foveal
and parafoveal form-priming paradigms
Jeremy Pacht1, W Schroyens2, & G d’Ydewalle2
Jeremy@psy.gla.ac.uk
1 University of Glasgow; 2 University of Leuven
In the past fifteen years, studies investigating the influence of a word’s orthographic hood have placed important constraints on models of visual word recognition One important finding,which is taken as evidence that word recognition involves competition between lexical candidates, is thatwords with a higher frequency neighbour elicit poorer performance than words having no such neigh-bours (e.g,., Grainger, O’Regan, Jacobs, & Segui, 1989) In their current instantiations, several influentialmodels of word recognition either do not or cannot simulate this effect (cf Forster, 1989; Sears, Hino, &Lupker, 1999; Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001) However, a number of investigatorshave argued that inhibitory neighbourhood frequency effects are unreliable or of limited relevance tovisual word recognition (e.g., Andrews, 1997; Forster & Shen, 1996; Siakaluk, Sears, & Lupker, 2002)
neighbour-We report a number of form-priming experiments in which the target is primed by a higher frequencyneighbour The results indicate that words may be inhibited by both foveal and parafoveal primes, andreinforce the claim that lexical resolution involves competitive processes
References
Andrews, S (1997) The effect of orthographic similarity on lexical retrieval: Resolving neighborhoodconflicts Psychonomic Bulletin & Review, 4, 439-461
Coltheart, M., Rastle, K., Perry, C., Langdon, R., & Ziegler, J (2001) DRC: A dual route cascaded model
of visual word recognition and reading aloud Psychological Review, 108, 204-256
Forster, K.I (1989) Basic issues in lexical processing In W Marslen-Wilson (Ed.), Lexical representationand process (pp 75-107) Cambridge, MA: MIT Press
Forster, K.I., & Shen, D (1996) No enemies in the neighborhood: Absence of inhibitory neighborhood fects in lexical decision and semantic categorization Journal of Experimental Psychology: Learning,Memory, & Cognition, 22, 696-713
ef-Grainger, J., O’Regan, J.K., Jacobs, A.M., & Segui, J (1989) On the role of competing word units in visualword recognition: The neighborhood frequency effect Perception & Psychophysics, 45, 189-195.Sears, C.R., Hino, Y., & Lupker, S.J (1999) Orthographic neighborhood effects in parallel distributedprocessing models Canadian Journal of Experimental Psychology, 53, 220-230
Siakaluk, P.D., Sears, C.R., & Lupker (2002) Orthographic neighborhood effects in lexical decision: Theeffects of nonword orthographic neighborhood size Journal of Experimental Psychology, 28, 661-681
Trang 22Prosodic phrasing in the RC-attachment ambiguity: Effects of
language, RC- length, and position
Eva M Fern´andez1; Dianne Bradley2; Jos´e Manuel Igoa3; Celia Teira3
eva fernandez@qc.edu, dbradley@gc.cuny.edu, josemanuel.igoa@uam.es, clteira@arrakis.es
1 Queens College & Graduate Center, CUNY
2 Graduate Center, CUNY
3 Universidad Aut´onoma de Madrid
Hemforth et al (submitted) explore cross-linguistic variation in attachment preference for a relative
clause (RC) modifying either of two sites in a complex NP RC’s length was varied with the usual result
(Fern´andez, 2003): More high attachment with longer RCs But in an important development, the host’s
syntactic position was simultaneously manipulated, and findings here involve a previously unreported
interaction with language: High attachment was more likely for object than subject position in Spanish
and German, but not English Hemforth et al.’s account appeals to two separate aspects of grammar
Language-sensitive position effects are attributed to differences in focus: Non-topic objects (focused in
the default, with focus falling on the head) will differ from subjects only in languages where preverbal
subjects are syntactically distinct topics, i.e., Spanish and German, but not English Length effects are
attributed to language-universal prosodic principles favoring separate phonological phrasing of heavy
constituents
However, the target construction in subject position is a super-heavy constituent; moreover the RC
modifier sets up a syntactic discontinuity at the matrix verb, regardless of RC’s attachment An NP][V
phrasing break is therefore likely, creating a global prosodic structure for subject-position sentences quite
different from that for object-position sentences There can be no guarantee that languages converge in
this prosodically intricate situation (e.g., Sandalo & Truckenbrodt, 2002), or that phrase-weight effects
local to RC can be considered in isolation (Carlson, Clifton & Frazier, 2001)
Prosodic theory being as yet a work in progress, our overt prosody study surveys patterns of
phono-logical phrasing as a function of RC-length and syntactic position, for American-English and
Castillian-Spanish Utterances containing the complex NP-RC construction were elicited using a variant of Bradley,
Fern´andez & Taylor’s (2003) “Post-to-Times” procedure, illustrated in (1) and (2); N=8 speakers per
lan-guage produced N=8x4 sentences drawn from Hemforth et al.s materials
Analyses of duration measures (indicating sites of consistent phrase-breaks, Bradley et al., 2003)
suggest that N2-lengthening before long RC is reliably modulated by syntactic position in English but
not Spanish The languages also differ with respect to the NP][V break when the target construction is a
subject, breaks being more substantial in Spanish than English, and this post- RC break likely tempers
interpretation of any N2][RC break (Carlson et al., 2001) We conclude that careful investigation of global
sentence prosody, beyond phrase-weight effects local to RC, is a prerequisite to evaluating Hemforth et
al.’s suggestion that discourse and prosody jointly influence RC- attachment
1 The guest impressed the brother of the bridegroom
(The brother of the bridegroom impressed the guest.)
Which bridegroom? The bridegroom who (often unknowingly) snores
2 El invitado impresion´o al hermano del novio
(El hermano del novio impresion´o al invitado.)
¿Qu´e novio? El novio que (a menudo inconscientemente) roncaba
References
Bradley, D., Fern´andez, E & Taylor, D (2003) Prosodic weight versus information load in the relativeclause attachment ambiguity Paper presented at the 16th Annual CUNY Conference on HumanSentence Processing, Cambridge MA
Carlson, K., Clifton, C., Jr., & Frazier, L (2001) Prosodic boundaries in adjunct attachment Journal ofMemory and Language, 45, 58-81
Fern´andez, E M (2003) Bilingual Sentence Processing: Relative Clause Attachment in English and ish Amsterdam: John Benjamins Publishers
Span-Hemforth, B., Fern´andez, S., Clifton, C Jr., Frazier, L., Konieczny, L., & Walter, M (submitted) Relativeclause attachment in German, English and Spanish: Effects of position and length
Sandalo, F & Truckenbrodt, H (2002) Some notes on phonological phrasing in Brazilian Portuguese.MIT Working Papers in Linguistics 43, 81-105
Trang 23Lexical effects in sentence processing: evidence from modifier
attachment in Greek and English
Despina Papadopoulou1, Theodore Marinis2& Leah Roberts3
despinap@otenet.gr, t.marinis@ucl.ac.uk, lroberts@essex.ac.uk
1 University of Crete; 2 University College London; 3 University of Essex
Recent sentence processing research has focused on relative clause (RC) attachment preferences in
ambiguous sentences such as (1), where the RC can refer either to the first NP the psychiatrist or to the
second one the singer
1 The man saw the psychiatrist of the singer who was playing chess
According to the thematic domain hypothesis (TDH) under the Construal account (Frazier &
Clifton, 1996), the thematic properties of prepositions linking the two potential host nouns determine
the preferred attachment site for the RC When the second NP is introduced by a thematic preposition,
such as with in (2), this signals the beginning of a new thematic domain Only NP2 lies within this
thematic domain, and thus, there is an NP2 attachment bias
2 The man saw the psychiatrist with the singer who was playing chess
The TDH has received support from several off- and on-line studies, which have compared
genitives/non-thematic prepositions to thematic prepositions in many languages, such as Italian (De
Vin-cenzi & Job, 1995), German (Hemforth et al., 1998) and Greek (Papadopoulou, 2002) However, all these
studies have used only one thematic preposition, with
The present study is the first looking at RC attachment preferences with several types of thematic
prepositions using both off- and on-line tasks and investigating RC attachment in two different languages,
English and Greek In order to test the predictions of the TDH and to explore the effects of lexical
prepositions in modifier attachment preferences in the two languages, we used two types of prepositions,
the preposition with and local prepositions in an off-line questionnaire and an on-line self-paced reading
experiment
40 English and 28 Greek subjects participated in our experiments The results of the off-line
exper-iment provided evidence in support of the TDH The subjects exhibited parallel attachment preferences
with all thematic prepositions in both English and Greek Conversely, the results of the on-line
experi-ment showed a different pattern While English subjects showed the same attachexperi-ment preferences for all
thematic prepositions, the Greek preposition me = with yielded different attachment preferences from
the locative preposition konda-se = near This shows that contrary to the predictions of the TDH not all
thematic prepositions yield the same attachment preferences We will discuss the implications of our
results for Construal and we will address the issue of cross-linguistic variation in sentence processing
References
De Vincenzi, M., & Job, R 1995 An investigation of late-closure: The role of syntax, thematic structure
and pragmatics in initial and final interpretation Journal of Experimental Psychology: Learning,
Memory and Cognition, 21: 1303-1321
Frazier, L., & Clifton, C 1996 Construal Cambridge, MA: MIT Press
Hemforth, B., Konieczny, L., Scheepers, C., & Strube, G 1998 Syntactic ambiguity resolution in German
Syntax and Semantics, 31: 293-309
Papadopoulou, D 2002 Cross-linguistic variation in sentence processing: evidence from relative clause
attachment preferences in Greek Unpublished PhD thesis: University of Essex, UK
Processing noncanonical constructions in free word order languages
Shravan Vasishthvasishth@coli.uni-sb.deUniversit¨at des Saarlandes
In free word order languages, noncanonical order is known to be harder to process than canonical der (e.g., Hy¨ona et al 1997), and this has often been attributed to factors such as increased argument-headdistance taxing limited working memory resources (e.g., Gibson 2000) However, Kaiser and Trueswell(2002) have recently shown that in Finnish, appropriate discourse context can neutralize the processingdifficulty associated with noncanonical order
or-Kaiser et al.’s experiments do not directly evaluate the role of working memory-based theories Thepresent research makes such a direct comparison and shows that resource limitations can indeed result inprocessing difficulty even when appropriate discourse context is present This result suggests that well-defined working memory constraints operate in conjunction with discourse-based contraints, and canoverride discourse-based facilitation under very specific conditions
Two self-paced reading studies using Hindi (a free word order language) were carried out to gate the effect of discourse context on processing ease Experiment 1 compared canonical order sentenceswith indirect-object fronted (IO-fronted) sentences with or without appropriate preceding context for thenoncanonical order; see (1) Similarly, Experiment 2 compared canonical order with direct-object fronted(DO-fronted) sentences, with a context manipulation similar to Experiment 1; see (2) The key differencebetween Experiment 1 and 2 was that in the former the fronted IO had only one intervening discoursereferent with respect to its canonical position, whereas in the latter experiment there were two interven-ing discourse referents between the fronted DO and its canonical position The number of interveningdiscourse referents was used to quantify argument-head distance (Gibson 2000)
investi-The prediction was that if increasing argument-head distance taxes working memory resources, thelonger distance in DO-fronted sentences would have a greater (and adverse) effect than discourse-basedfacilitation, whereas in IO-fronted sentences discourse-based facilitation would not be overridden (or not
to as great an extent)
This prediction was borne out: in canonical versus DO-fronted sentences, even with appropriatediscourse context the RTs at the verb were longer in the DO-fronted case, whereas in canonical versusIO-fronted sentences, appropriate context neutralized RTs at the verb
In sum, independent of discourse-related factors, the extent to which working memory resourcesare taxed appears to be a critical factor: in free word order languages if an argument is sufficiently distantfrom the verb selecting for it, the presence of appropriate discourse context ceases to facilitate processing.Examples
1 a Rita-ne Ravi-ko kyaa kahaa? Rita-ne Ravi-ko kitaab khariid-neko kahaaRita-erg Ravi-dat what said? Rita-erg Ravi-dat book buy-inf told
‘What did Rita say to Ravi? Rita told Ravi to buy a book.’
b Ravi-ko Rita-ne kyaa kahaa? Ravi-ko Rita-ne kitaab khariid-neko kahaaRavi-dat Rita-erg what said? Ravi-dat Rita-erg book buy-inf told
‘What did Rita say to Ravi? Rita told Ravi to buy a book.’
c Kyaa hua? Rita-ne Ravi-ko kitaab khariid-neko kahaaWhat happened? Rita-erg Ravi-dat book buy-inf told
‘What happened? Rita told Ravi to buy a book.’
d Kyaa hua? Ravi-ko Rita-ne kitaab khariid-neko kahaaWhat happened? Ravi-dat Rita-erg book buy-inf told
Trang 242 a Rita-ne Ravi-ko kyaa kahaa? Rita-ne Ravi-ko kitaab khariid-neko kahaa
Rita-erg Ravi-dat what said? Rita-erg Ravi-dat book buy-inf told
‘What did Rita say to Ravi? Rita told Ravi to buy a book.’
b Kitaab ka Rita-ne kyaa kiyaa? kitaab Rita-ne Ravi-ko khariid-neko kahii
book gen Rita-erg what did? book Rita-erg Ravi-dat buy-inf told
‘What did Rita do about the book? Rita told Ravi to buy the book.’
c Kyaa hua? Rita-ne Ravi-ko kitaab khariid-neko kahaa
What happened? Rita-erg Ravi-dat book buy-inf told
‘What happened? Rita told Ravi to buy a book.’
d Kyaa hua? kitaab Rita-ne Ravi-ko khariid-neko kahii
What happened? book Rita-erg Ravi-dat buy-inf told
‘What happened? Rita told Ravi to buy a book.’
References
Gibson, E 2000 Dependency locality theory: A distance-dased theory of linguistic complexity Image,
Language, brain: Papers from the First Mind Articulation Project Symposium MIT Press
Jukka Hy¨ona and Heli Hujanen 1997 Effects of Case Marking and Word Order on Sentence Parsing in
Finnish: An Eye-Fixation Study, Vol 50A(4), pp 841-858
Elsi Kaiser and John C Trueswell 2002 A new ’look’ in the processing of noncanonical word orders
CUNY conference, New York City
Poster Session 1: Monday 25th August
Trang 25Evaluating Probabilistic Models of Human Parsing
Frank Kellerkeller@inf.ed.ac.ukSchool of InformaticsUniversity of Edinburgh
A number of probabilistic models of human parsing have been proposed in the literature (Crocker
& Brants 2000; Hale 2003; Jurafsky 1996) All of these models are based on a probabilistic context-free
grammar (PCFG) trained on an annotated corpus The predictions of the PCFG are then evaluated against
a set of psycholinguistically relevant test sentences (typically garden path sentences)
There are two shortcomings of this approach: (1) It is well known from computational linguistics
that the type of grammar used (e.g., lexicalized or unlexicalized, Charniak 1997) can have a large effect
on parsing performance However, this effect is not explored in the psycholinguistic modeling literature
(2) The evaluation of the models of Crocker & Brants (2000), Hale (2003), and Jurafsky (1996) is
ba-sically anecdotal: only a handful of carefully selected example sentences are tested against the model,
and there is no quantitative evaluation of model predictions This makes it very had to compare the
performance of different models
In this paper, we address both of these shortcomings We evaluate the performance of three
PCFG-based parsing models that differ in the type of linguistic information they incorporate The Baseline
Model is based on a standard unlexicalized PCFG The Lexicalized Model uses a lexicalized PCFG
as proposed by Charniak (1997); this model incorporates lexical information and hence approximates
selectional restrictions, semantic plausibility, and morphological dependencies The Functional Model
uses a PCFG augmented with information about grammatical functions (subject, object, etc.) and can
therefore capture syntactic dependencies
All three models are trained on Negra, a syntactically annotated corpus of German newspaper text
(Skut et al 1997) and tested against two large magnitude estimation datasets (128 and 192 sentences)
dealing with word order variation in subordinate clauses in German (Keller 2000a,b) Evaluation is
per-formed by correlating the probabilities that a model assigns to the sentences in the test set with the
magnitude estimation score obtained experimentally
Significant correlations between model predictions and experimental data are obtained for all three
models There are, however, important differences in modeling performance: the lexicalized model
achieves a better fit with the experimental than the baseline model This indicates that the lexical
informa-tion that is approximated by the lexical model, viz., morphological informainforma-tion and semantic plausibility,
plays an important role in human parsing The functional model performs poorly, it fails to outperform
the baseline model This could be due to sparse data: there are not enough corpus examples available to
reliably estimate the parameters of the functional model
Magnitude estimation data are offline data that only represent a rough approximation of processing
difficulty We are therefore currently extending our models to deal with online data such as eye-tracking
and self-paced reading data Preliminary research indicates that our modeling results generalize to these
data types
References
Charniak, E (1997) Statistical parsing with a context-free grammar and word statistics In Proceedings
of the 14th National Conference on Artificial Intelligence, (pp 598-603), Cambridge, MA AAAI
Keller, F (2000a) Evaluating competition-based models of word order In L R Gleitman, & A K Joshi(eds.), Proceedings of the 22nd Annual Conference of the Cognitive Science Society, (pp 747-752),Mahwah, NJ Lawrence Erlbaum Associates
Keller, F (2000b) Gradience in Grammar: Experimental and Computational Aspects of Degrees of maticality Ph.D thesis, University of Edinburgh
Gram-Skut, W., Krenn, B., Brants, T., & Uszkoreit, H (1997) An annotation scheme for free word order guages In Proceedings of the 5th Conference on Applied Natural Language Processing, Washington,DC
Trang 26lan-Event-types in pronoun resolution
Barbara Hemforth1, Lars Konieczny2, and Joel Pynte1
barbara.hemforth@lpl.univ-aix.fr
1 University of Provence, 2 University of Freiburg
Based on questionnaire data as well as eyetracking experiments, Hemforth and Konieczny (2002)
argued that preferences in pronoun resolution in sentences like (1) respect the information structure of
the sentence, i.e., pronouns tend to prefer antecedents in topic-position (the actress in 1)
(1) Die Schauspielerin faszinierte den Regisseur, als sie/er hinter der Buehne stand
The actress[fem] fascinated the director[masc] when she/he was standing behind the stage
More recent self-paced reading experiments in French and German show, however, that this
pref-erence is dependent on the event-type described in the sentence Closely matched self-paced reading
experiments were run in German and French varying the position of the antecedent of the pronoun as
well as the connective (2 to 5)
(2) Le journaliste interviewe la politicienne apr`es quil/elle est sorti de la salle de conf´erence
Der Journalist interviewt die Politikerin, nachdem er/sie den Konferenzsaal verlassen hat
The journalist[masc] interviews the politician[fem] after he/she has left the conference room
(3) Le journaliste interviewe la politicienne avant quil/elle est sorti de la salle de conf´erence
Der Journalist interviewt die Politikerin, bevor er/sie den Konferenzsaal verlassen hat
The journalist[masc] interviews the politician[fem] before he/she has left the conference room
(4) Le journaliste interviewe la politicienne pendant quil/elle sort de la salle de conf´erence
Der Journalist interviewt die Politikerin, waehrend er/sie den Konferenzsaal verlaesst
The journalist[masc] interviews the politician[fem] while he/she is leaving the conference room
(5) Le journaliste interviewe la politicienne quand il/elle sort de la salle de conf´erence
Der Journalist interviewt die Politikerin, als er/sie den Konferenzsaal verlaesst
The journalist[masc] interviews the politician[fem] when he/she is leaving the conference room
Across connectives as well as across languages the preference for topical antecedents was lost or
even turned into a preference for the non-topical second NP
The major difference between the earlier experiments on German and the cross-linguistic
experi-ments to be discussed here is the event-type of the verbs used in the materials Whereas Hemforth and
Konieczny used only psych-verbs to allow for variable word ordering, all verbs in the French/German
ex-periments were action verbs Following Stevenson et al (1994, 2000) action verbs focus the patient role
(i.e., NP2), whereas no particular focus is imposed by psych-verbs Therefore a preference for topical
an-tecedents only shows through when materials containing psych-verbs are used Experiments comparing
action verbs and psych-verbs more directly in German and French are currently underway
The dynamics of plural-markings in agreement errors in production
Lars Konieczny1, Sarah Schimke1and Barbara Hemforth2
lars@cognition.iig.uni-freiburg.de
1 University of Freiburg; 2 University of Provence
We will present an ACT-R-based model of verb-processing that captures the known evidence onagreement errors in production:
• Local Plural error after complex Subject NPs (Modifier attraction, “feature percolation” (cf.Vigliocco & Nicol, 1998 )
• General singular errors (singular verb produced when Subject is plural, regardless of local noun,
cf Hemforth and Konieczny, 2002)
• Reduced singular errors when Object is Plural in S-O-V constructions (“feature reactivation”,Hemforth and Konieczny, 2002)
• Plural attraction from Objects in S-O-V constructions under time pressure (Object attraction ,Hartsuiker, 2002)
• The lack of Plural attraction from Object-NPs under no-pressure or no-load conditions (Hemforth
Trang 27Relative Clause Attachment in French: an ERP study
J Pynte1, C Portes1, P Holcomb2, A Di Cristo1
Email: pynte@up.univ-aix.fr
1 University of Provence, 2 Tufts University
Relative Clause Attachment preferences have been shown to be influenced by semantic, pragmatic
and prosodic factors (for evidence concerning the role of prosody and animacy, see Schafer et al., 1996
and Mak et al., 2002, respectively) This does not mean that structural principles play no role in
com-prehension For example, Late Closure could take some part in the initial structuring of the utterance
Evidence of initial low-attachment (and final high-attachment) preference was found in an ERP
experi-ment conducted in French 16 native speakers of French were orally presented with 96 sentences of the
type “n0+V+n1+of+n2+RC” (Example 1) They were required to perform an acceptability judgement at
the end of each trial Three factors were manipulated, namely (i) Antecedent animacy: for half of the
sen-tences, n1 denoted a human entity and n2 an inanimate entity, whereas for the other half, the opposite was
true; (ii) Prosody: a pitch accent was or was not present on the last syllable of n1; and (iii) Attachment:
disambiguation either forced low or high attachment A N400 effect was expected in the non-preferred
disambiguation condition Suppose that “keeper” is interpreted as the antecedent of the relative pronoun,
whereas the disambiguating word “garden” is presented This leads to the incongruous statement that the
keeper “had been described as a garden”, which would elicit a relatively important N400 component
(1) Les badauds ont rencontr´e le surveillant du parc qui avait ´et´e d´ecrit comme un gardien/jardin
agr´eable
(The idler met the keeper of the park who/which had been described as a guard/garden agreeable)
The analysis of acceptability judgements revealed a general preference for attaching to n1 on the
one hand (F(1,12)=25.11), and to the noun denoting a human entity on the other hand (F(1,12)=19.68, no
effect of prosody) (The same pattern of preferences was independently found with a group of 32 judges)
A preference for attaching to the noun denoting a human entity was also found on line (less negativity
during the presentation of the disambiguating word) However the N400 component was smaller, overall,
when disambiguation forced n2 attachment (F(1,12) = 7.57, p < 05) The presence of a pitch accent on
n1 did hindered n2 attachment, but only in the n1-human condition (F(1,12) = 4.26 for the three-way
interaction) The same pattern of results was found, whether “less acceptable” trials were included in the
analysis or not The implications of these results for current models of parsing will be discussed
References
Mak, W M, Vonk, W., Schriefers, H (2002) The influence of animacy on relative clause processing
Journal of Memory and Language, 47, 50-68
Schafer, A., Carter, J., Clifton, C Jr., and Frazier, L (1996) Focus on relative clause construal Language
and Cognitive Processes, 11, 135-163
Pre-verb thematic role assignment and -revision in German verb-final constructions: Counter-evidence from eye-movements in reading
Christoph Scheepers1and Edith Klee2
c.scheepers@dundee.ac.uk edith@coli.uni-sb.de
1 University of Dundee, 2 University of Saarland
In an ERP-study, Bornkessel, Schlesewsky, & Friederici (in press) found that dative-nominativestructures (1b), when compared to nominative-dative structures (1a), elicit an early (200ms post-onset)positivity in the NP2-region, but no effect in the NP1-region They attribute this to pre-verb thematicrole assignment and -revision: just as nominative-NPs, animate dative-NPs in clause-initial position aretentatively interpreted as proto-agents - an analysis that has to be revised as soon as a ‘better’ proto-agent
is available (the nominative-NP2 in 1b)
(1) a ., daß der Pfarrer damals dem Bischof geholfen hatte
, that the priest [nom] then the bishop [dat] helped had
b ., daß dem Pfarrer damals der Bischof geholfen hatte
, that the priest [dat] then the bishop [nom] helped had
c ., daß der Pfarrer damals vom Bischof unterst¨utzt wurde
, that the priest [nom] then by-the bishop supported was
d ., daß dem Pfarrer damals vom Bischof geholfen wurde
, that the priest [dat] then by-the bishop helped was
The present eye-tracking study aimed at replicating these results and testing further predictions
of the model Besides the active voice conditions in (1a,b), our experiment included the passive voiceconditions in (1c,d) Assuming that clause-initial nominative or dative NPs are interpreted as proto-agentsupon first encounter, readers should experience processing difficulty when they reach the by-phrase in(1c,d), since by-phrases are likely to indicate actual agents (i.e., better proto-agents than either of the
previous NPs) note that in nominative-first passives (1c), the by-phrase must be interpreted as an agent.
However, none of the predictions were supported Contrary to Bornkessel et al (in press), we foundimmediate processing difficulty (in first-fixation, first-pass, and regression-path time the latter also takesfist-pass regressions from a region into account) for clause-initial dative-NPs (‘dem Pfarrer’, 1b,d) ascompared to clause-initial nominative-NPs (‘der Pfarrer’, 1a,c), but no reliable effect whatsoever in theNP2/by-phrase region We conclude (a) that the ‘early positivity’ around NP2 in the ERP experimentmight actually have been triggered by NP1 and was just carried over to NP2, and (b) that our ownfindings support a purely syntactic nominative-first preference Thus, for reading isolated sentences ourdata do not support the claim that thematic role assignments and -revisions take place before the verb isencountered
References
Bornkessel, I., Schlesewsky, M., & Friederici, A D (in press) Eliciting thematic reanalysis effects: Therole of syntax-independent information during parsing Language and Cognitive Processes
Trang 28Event structure effects on garden pathing
Erin L O’Bryan1, Raffaella Folli2, Heidi Harley1and Thomas G Bever1
obryan@u.arizona.edu, rf250@cam.ac.uk, hharley@u.arizona.edu, tgb@u.arizona.edu
1 University of Arizona, 2 University of Cambridge
We present a reading and an auditory study, showing that event structure information, specifically
telicity, is accessible during comprehension Cross-linguistic research suggests that verbs denoting telic
events (events that progress towards an endpoint), even intransitive ones, have underlying direct objects
(Tenny, 1992) The information that an object is required may lead the parser to leave open the possibility
that the initial noun phrase is that object This led Sanz (2000) to predict that garden path effects will be
smaller in reduced relative sentences with telic verbs compared to those with atelic verbs O’Bryan et al
(2002, 2003) found support for this prediction in reanalyses of prior self-paced reading experiments
To directly test the prediction that less garden pathing occurs in reduced relatives with telic verbs,
we conducted a reduced relatives reading experiment in which we balanced the number of verbs in 4
categories, crossing telicity and transitivity The transitivity categorization was determined by a
question-naire testing each verbs naturalness in different argument structure frames, and the telicity categorization
by standard event structure tests In (1), the verb conditions are optionally transitive atelic, optionally
transitive telic, obligatorily transitive atelic, and obligatorily transitive telic
(1) The actress (that was) sketched/awakened/described/spotted by the writer left in a hurry
We used the word maze paradigm (Freedman & Forster, 1985) because it provides RTs at each word
without the problem of spillover found in self-paced reading In this task, subjects inspect successive
word pairs, in which only 1 word forms a grammatical continuation of a sentence The subject’s task is
to pick that word The word-choice RTs revealed less garden pathing on ”by” with telic than atelic verbs
(p < 01) Less garden pathing for obligatorily transitive verbs was found only later, e.g on the main
verb
In a separate auditory study with similar materials, we confirmed the role of telicity, using a
voice-change detection paradigm (Townsend & Bever, 1991), in which the voice voice-change occurred on the noun
of the embedded relative clause (‘writer’ in (1)) The results showed fewer detection errors due to reduced
relative clauses with telic verbs, again suggesting less garden pathing (Verb transitivity had a weaker and
opposite effect)
We conclude that event structure information is accessible during on-line parsing separately from
and prior to argument structure This suggests further that event structure is not derived from argument
structure
References
Freedman, S E & K I Forster (1985) The psychological status of overgenerated sentences Cognition
19, 101-131
O’Bryan, E L., R Folli, H Harley, & T G Bever (2003) Event structure is accessed immediately during
comprehension Paper presented at the Linguistics Society of America conference, Atlanta, Georgia
O’Bryan, E L., R Folli, H Harley, C Clarke, D J Townsend, & T G Bever (2002) The role of event
structure in language comprehension Poster presented at the AMLaP conference, Tenerife, Spain
Sanz, M (2000) Events and Predication: A New Approach to Syntactic Processing in English and Spanish
Amsterdam: John Benjamins
Tenny, C (1992) The Aspectual Interface hypothesis In I A Sag & A Szabolcsi (eds.), Lexical Matters
Palo Alto, CA: Stanford University
Townsend, D J & T G Bever (1991) The use of higher-level constraints in monitoring for a change
in speaker demonstrates functionally distinct levels of representation in discourse comprehension
Language and Cognitive Processes 6, 49-77
Three way attraction effects in Slovenian
Annabel Harrison1, Holly Branigan1, Rob Hartsuiker2& Martin Pickering1
annabelh@cogsci.ed.ac.uk, holly.branigan@ed.ac.uk, robert.hartsuiker@rug.ac.be,
martin.pickering@ed.ac.uk
1 University of Edinburgh, 2 University of Ghent
Attraction effects are well established with plural local nouns (e.g., Bock & Miller, 1991) and haverecently been observed with singular local nouns (Haskell & Bock, 2003)
Agreement processes involving more complex number systems have long been the subject of bate in the linguistics literature (see Corbett, 2000), but psycholinguists have thus far concentrated onattraction effects in two-way systems
de-We report a study that employed the three-way number distinction in Slovenian to address the issue
of markedness, namely, whether there is a binary distinction between marked plurals and unmarkedsingulars as claimed by Eberhard (1997) for English
Eberhard’s (1997) model would predict a marked form and an unmarked form, but could not explaindifferences between the two unmarked forms If Slovenian follows the same pattern as English, wewould expect the singular to be the unmarked default and thus susceptible to agreement errors yet notcausing them; conversely, the dual and plural would form the marked class, and would cause errors onthe singular, but not be prone to errors after a singular local noun Dual and plural marking in the sameitem would be in equal competition
If the model were to incorporate a third level of markedness, then differences between the dual andplural could be explained, provided that the singular was unmarked According to Corbett (2000), wewould expect the markedness to be ordered: singular < plural < dual
90 native Slovenian speakers performed a sentence completion task involving a preamble (a plex NP containing a head noun and a postmodifying relative clause), and a verb independently rated asmore plausible with the head than the local noun The number values of the head and local noun weremanipulated, yielding 9 different conditions
com1 Bik (Bika/i), ki (sta/so) ga (ju/jih) je zabodel(zabodla/li) ponosen(ponosna/ni) matador (/ja/ji) raniti se
-Bull(Bulls-d/p) that AUX-d/p OBJ-s d/p AUX-s stab-s (stab-d/p) proud-s (proud-d/p) matador(matadors-d/p) - injure REFLEXIVE
“A bull (bulls) that a proud matador (proud matadors) stabbed - injure oneself”
Of the correctly produced preambles followed by a number-inflected verb, the agreement errors(number value did not equal that of head noun) and attraction errors (number value equalled that of localnoun) are shown below:
% agreement errors % attraction errors
Trang 29-We see agreement effects with a singular local noun and smaller effects after a plural local noun,
as well as larger effects after a dual local noun These results are discussed in terms of implications for
models of agreement production
References
Bock, J K & Miller, C A (1991) Broken agreement Cognitive Psychology, 23, 45-93
Corbett, G G (2000) Number Cambridge: Cambridge University Press
Eberhard, K M (1997) The marked effect of number on subject-verb agreement Journal of Memory and
Language, 36, 147-164
Haskell, T & Bock, J K (2003) Singular attraction in subject-verb agreement Paper presented at the 16th
CUNY Conference on Human Sentence Processing 2003
Does the Asymmetric Neighbourhood Effect interact with handedness
in English Readers?
Alexandra McCauley, Richard Shillcockalexmcca@cogsci.ed.ac.uk, rcs@cogsci.ed.ac.ukDivision of Informatics, University of Edinburgh
In orthographic lexical decision tasks facilitation is usually seen for words with many neighbours(N) This is termed the neighbourhood effect (NE) N is defined as the number of same length wordsdiffering from a target word by one letter A written word may activate lexical entries for such “similar”words, influencing recognition speed The NE has also been investigated in a lateralised lexical deci-sion task Lavidor and Ellis (2002) report an asymmetric NE (aNE) in which a NE is seen in the righthemisphere (RH) but not in the left hemisphere (LH) in right-handed native English speakers
In the split-fovea model, visual word recognition is conditioned by the precise splitting of the fovealimage about the fixation point, with letters to the left and right of fixation projected contralaterally (Shill-cock et al 2000) The two halves of a word have different statistical profiles, with more informativeletter sequences at word beginnings (Yannakoudakis & Hutton, 1992; McCauley, 2002) RH processingreflects processing of these initial, more informative, letters It seems adaptive for the RH to activatelead-N, neighbours of the initial letters of a word, while this is not so for end-N in the LH Manipulat-ing lead and end-N values of words whilst keeping the N value of the whole word constant has shownfacilitation for words with many lead-N but no effect for words with many end-N (Lavidor and Walsh,2003)
If the RH applies this strategy when performing lexical decision on the whole word, we wouldexpect the aNE; the RH continues to activate N while the left hemisphere does not To begin to test thishypothesis I report a study which aims to evaluate differences in the expression of the aNE betweenleft and right-handers A proportion of left-handers are known to differ in language lateralisation and inhemispheric asymmetries for lexical processing However, left-handers share the same reading exposure
as right-handers Therefore, our strongest hypothesis states that left-handers should not differ in theirexpression of the aNE from right-handers
Participants are presented with a battery of tests including the Edinburgh Handedness Inventoryand a lateralised lexical decision task manipulating word neighbourhood The current methodology wasable to replicate established effects It also suggests that left-handers do differ from right-handers inthe effect of N on lexical processing in the two hemispheres, despite consistency of reading exposure.This provides a further constraint for the split-fovea model Similar investigations have been undertakenwith Hebrew Hebrew words also have more informative word beginnings, but this time on the right ofthe word (McCauley, AMLaP, 2003) I comment finally on a comparison study being run with left andright-handed Hebrew readers
Trang 30Shillcock, R., Ellison, T M & Monaghan, P (2000) Eye-Fixation Behavior, Lexical Storage, and Visual
Word Recognition in a Split Processing Model Psychological Review, 107, 824-851
Yannakoudakis, E J., & Hutton, P J (1992) An assessment of N-phoneme statistics in phoneme guessing
algorithms which aim to incorporate phonotactic constraints Speech Communications, 11, 581-602
Detecting text changes as a function of load and extent of change:
what’s the mechanism?
Alison Sanford1, Jason Bohan2, Anthony Sanford2and Jo MolleAlison.sanford@strath.ac.uk, jason@psy.gla.ac.uk, tony@psy.gla.ac.uk, jomolle@hotmail.com
1 University of Strathclyde, 2 University of Glasgow
This paper reports a series of experiments that explore the effects of cognitive load on semanticprocessing A recently developed change-detection paradigm was used, in which participants read a texttwice: on the second presentation, a change may or may not occur The index of performance is theproportion of changes detected, and it has been suggested that this reflects the degree of processingafforded the aspect of the text that changes (Sanford & Sturt, 2002)
The technique is potentially useful for investigating the dependence of extent of processing oncognitive load imposed during reading In a short series of experiments to be reported, cognitive loadwas manipulated by the use of embedded sentences (following Gibson’s approach), in which load wasmanipulated by using object-extracted centre embedded clauses (high load), or subject extracted clauses(low load): high load results in fewer detections In addition, experiments in which load was manipulated
by using first person pronouns or full noun phrases in relative clauses were carried out (Warren & Gibson,2002) Under high load (full NP), fewer detections were made on the main verbs of the target sentencethan under low load (pronoun) conditions Testing for load effects at other points in the sentence showed
no effect
Apart from presenting new data about load effects, the interest in the paper will focus on the jointeffects of semantic distance of change and load effects on change According to the granularity of rep-resentation account (Hobbs, 1985; Sanford et al., submitted), there should be an interaction of these twovariables, such as has been found for the effects of discourse focus on change detection In the stud-ies reported here, additive effects were found, suggesting that a different mechanism may underlie therelationship of load to detection Among the possibilities is the idea that load results in an increasedpossibility of failing to encode the clause in question: when the clause is encoded, however, the chances
of recognising a change depend upon semantic distance On this account, semantic distance may be arecognition memory effect These issues will be discussed
References
Gibson, E (1998) Linguistic complexity: locality of syntactic dependencies Cognition, 68, 1-76
Hobbs, J.R (1985) Granularity International Joint Conference on Artificial Intelligence 1985: 432-435.Sanford, A J., Sturt, P., Stewart, A., & Dawydiak, E (submitted) Linguistic Focus and Good-Enoughrepresentations: an application of the change-detection paradigm Submitted to Psychonomic Bulletinand Review
Sanford, A.J., & Sturt, P (2002) Depth of processing in language comprehension: not noticing the dence Trends in Cognitive Sciences, 6 (9), 382-386
evi-Warren, T & Gibson, E (2002) The influence of referential processing on sentence complexity Cognition,
85, 79-112
Trang 31Effects of prosody on the resolution of word-order ambiguities
Andrea Weber, Martine Grice, Matthew Crockeraweber@coli.uni-sb.de, mgrice@coli.uni-sb.de, crocker@coli.uni.sb.de
University of Saarland
In German, noun phrases (NPs) can be ambiguously case-marked as nominative (Subject,
typi-cally Agent) or accusative (Object, typitypi-cally Patient) Furthermore, both Agent and Patient can appear
sentence-initially, but Agent-first is canonical A recent visual-world study showed that, in the absence
of clear case marking, German listeners preferably interpret sentence-initial NPs as Agents (Kn¨oferle,
Crocker, Scheepers, & Pickering, 2001): Following case-ambiguous first NPs, anticipatory eye
move-ments to the picture of a Patient were observed, well before a disambiguating second NP It has already
been shown hat prosody can influence syntactic attachment ambiguities (see e.g., Kjelgaard & Speer,
1999) The present study investigated whether prosody can also manipulate the interpretation of
word-order ambiguities, using sentences with case-ambiguous first NPs and post-verbal second NPs with
un-ambiguous accusative (1) or nominative (2) case marking
(1) Die Katze(L∗+H)jagt gleich den Vogel
The cat (NOM, ambiguous) chases in-a-moment the bird (ACC)
(2) Die Katze(L+H∗)jagt gleich der Hund
The cat (ACC, ambiguous) chases in-a-moment the dog (NOM)
For the Agent-first reading (1) our speaker placed a low pitch accent (L*+H) on the first NP These
NPs were considered unmarked and expected to indicate canonical Agent-first sentences For the
Patient-first reading (2) she instead used a rising pitch accent (L+H*) Those NPs were considered marked and
expected to indicate non-canonical Patient-first sentences Recorded sentences were presented along with
scenes portraying the ambiguous character (cat), the Patient (bird), the Agent (dog), and a distractor
object Actions were not displayed, and the ambiguous character was equally likely as Agent or Patient
Fewer anticipatory looks to the Patient were predicted for (2) than for (1) Indeed, before the onset of
the second NP, the Patient was fixated more often than the Agent when the first NP was L*+H (1),
but not when it was L+H* (2) Thus, the interpretation of word-order ambiguities was modulated by
prosody However, in (2), prosody was not sufficient to reverse the preference for the canonical
Agent-first structure Interestingly, the effect of prosody shifted in time during the experiment In the Agent-first half,
sentence type (Agent-first, Patient-first) interacted with character (Patient, Agent) during the adverb
During the verb more looks to the Patient were found for both sentence types In the second half, sentence
type already interacted with character during the verb More looks to the Patient were observed for
Agent-first sentences only This suggests that listeners adapted to prosodic cues Importantly, however, in both
halves prosodic effects were found prior to the second NP In sum, we show that prosody can manipulate
word-order ambiguities: In the absence of clear case marking, prosodic cues were integrated rapidly
enough to affect listeners interpretation before disambiguating acoustic information was available
References
Kn¨oferle, P., Crocker, M., Scheepers, C., & Pickering, M (2002) Anticipatory eye movements in initially
ambiguous sentences: Theres more to it than meets the eye AMLaP conference, Tenerife, Spain
Kjelgaard, M., & Speer, S (1999) Prosodic facilitation and interference in the resolution of temporary
syntactic closure ambiguity Journal of Memory and Language, 40, 153-194
Controlling attention and structure in dialogue: the interlocutor v the
clock
E G Bard, A H Anderson, M Flecha-Garcia, D Kenicer, J Mullin, H Nicholson, L Smallwood, &
Y Chenellen@ling.ed.ac.uk, monitor@mcg.gla.ac.uk
This work contrasts two views of the process of participating in dialogue One holds that tal, interactive adjustment to the interlocutor is of primary importance regardless of cost The other holdsthat cognitive cost is critcal, with on-line adjustments less favoured than less costly global settings Anexperiment using the Map Task, a route communication task (Brown et al, 1983; Anderson et al., 1991)tests these proposals Speakers produced route descriptions for absent listeners in a factorial design con-trasting levels of time pressure (1 minute v unlimited) and feedback (some, none) Time pressure waspresented as a single global indication of the time limit Feedback was available solely via a square mov-ing across the map and purportedly representing the listener’s eye fixations, but actually directed by anexperimenter toward pre-arranged sequences of correct and incorrect landmarks Twenty-four subjectsproduced monologues in all 4 conditions
incremen-The interactionist position predicts strong effects of the feedback, while the cost-based approachpredicts strong effects of time pressure Measures of listeners’ attention (genuine eye fixations), of con-versational structure (conversational transactions), and length in words all show the same effects: robustchanges with time pressure (all F-values at p < 05) and little effect of feedback We discuss these effectswith respect to a model of the speaker’s priorities in spoken dialogue
Trang 32Accents, structure, and the interpretation of gapping sentences
Katy Carlson , Chris Kennedy, & Michael Walsh Dickeykatyc@northwestern.edu, kennedy@northwestern.edu, dickey@ling.nwu.edu
Northwestern University
Gapping sentences with a subject/object ambiguity (1) have a very strong object bias
(1) Bob insulted the guests during dinner and Sam during the dance
In auditory questionnaires (Carlson 2001a, 2002), 30% of responses interpreted Sam as a subject
(Sam also insulted guests) when the object (the guests) and Sam were accented, vs 46% when the subject
(Bob) and Sam received accents If the object reading involves a structurally simpler analysis than the
subject reading, then the strong object preference and small accentuation effect are natural Carlson
(2001a) suggested that the object analysis of (1) coordinates VP-level constituents, while the subject
reading coordinates entire sentences The present research supports this theory by showing that gapping
sentences which disallow VP coordination have more mobile preferences, as other ellipsis sentences
without structural biases do The remaining object bias of a range of ellipses is traced to information
structure (Frazier & Clifton 1998)
The current experiments studied gapping sentences with preposed PPs:
(2) At Marshall Field’s, Melissa saw a classmate, and at J.C Penney’s, Sabrina
Preposed locative and temporal PPs (e.g., at J.C Penney’s) adjoin to clausal sentence projections
like vP or IP (Reinhart 1983) Assuming Johnson’s (1996, 2000) syntax for gapping, the remnant
(Sab-rina) following a preposed PP also adjoins to vP, whether it came from subject or object position in the
VP, after which the whole VP is deleted Hence the post-conjunction material in (2) has the structure in
(3) on either interpretation:
(3) and [vP at Penney’s [vP Sabrina [vP (v VP: deleted) ]]]
In a written questionnaire, sentences like (2) received 36% subject responses (Sabrina also saw a
classmate) In an auditory questionnaire, accenting the first-clause subject (Melissa) and the remnant
produced 60% subject responses, while object (classmate) and remnant accents produced 15% subject
responses Thus with a preposed PP, gapping sentences are much more ambiguous and responsive to
accent placement than without
Comparative ellipses (4) are similarly ambiguous (Carlson 2001b):
(4) Tasha/He called him/Bella more often than Sonya
With subject parallelism (Tasha/him/Sonya), such comparatives received 68% subject responses;
with object parallelism (He/Bella/Sonya), they received 18% subject responses Like (2), these sentences
arguably do not have different syntactic structures on subject vs object readings (Kennedy 1998), yet
they show a general object bias We follow F&C 1998 in suggesting that this bias, and that of sentences
like (2), reflect a tendency to assign focus low in the VP
Self-corrections in speech: Evidence against Levelt’s Main Interruption
Rule
Ciara Catchpole1, Robert Hartsuiker2and Martin Pickering1
ciaracatchpole@angelfire.com, robert.hartsuiker@rug.ac.be, martin.pickering@ed.ac.uk
1 University of Edinburgh, 2 University of Ghent
The analysis of speech errors can tell us a great deal about the architecture of the language system,and in particular, the role played by monitoring devices in production In this study, we elicited lemma-substitution speech errors by means of a picture-naming task (inspired by Van Wijk and Kempen, 1987)where on a small number of trials the first picture (the interrupted stimulus, or IS) would change toanother picture (the corrected stimulus, or CS), after 300ms, at which point the subject was still namingthe IS The CS picture varied in terms of degradedness, a visual characteristic that reportedly affectsnaming latency (Meyer, Sleiderink and Levelt, 1998, whose results were also replicated as part of thisexperiment) The subjects were told that the second (CS) picture was always more important, and thatthey should correct themselves as quickly as possible if the picture changed Many subjects did not self-interrupt at all, and instead completed the first picture-name before beginning the second These fullrepair responses were removed from the data set for separate analysis, and the additional elimination oferrors left only the self-interruptions remaining, comprising approximately 25% of the data The timings
of these self-interruptions and their subsequent repairs were measured using manual waveform analysis.This use of degraded pictures in the CS set allowed us to test Levelt’s (1983) Main Interruption Rule,which states that a person’s flow of speech will be stopped immediately upon detecting a self-producederror If this were the case, then the interruption point of a person’s speech should be unaffected bycharacteristics of the repair Thus, it should not be affected by any attributes of the succeeding picture(the CS) This theory was not supported It was found that the time to interrupt the IS increased when the
CS was degraded This demonstrates that the flow of speech could not have been stopped immediately
We therefore rejected Levelt’s Main Interruption Rule, and concluded that it is more likely that when aspeech error is detected, the interruption and the repair are planned in parallel (Hartsuiker & Kolk, 2001).Importantly, the interruption is delayed until the repair has been planned to a certain extent It is likelythat this would occur in an effort to maintain the fluency of speech, by lessening the interval between theinterruption of the erroneous word and the onset of the repair
References
Hartsuiker, R J., & Kolk, H H J (2001) Error monitoring in speech production: A computational test ofthe perceptual loop theory Cognitive Psychology, 42, 113-157
Levelt, W J M (1983) Monitoring and self-repair in speech Cognition, 14, 41 104
Meyer, A S., Sleiderink, A M., & Levelt, W J M (1998) Viewing and naming objects: eye movementsduring noun phrase production Cognition, 66, B25-B33
Van Wijk, C., & Kempen, G (1987) A dual system for producing self-repairs in spontaneous speech:Evidence from experimentally elicited corrections Cognitive Psychology, 19, 403-440
Trang 33The Nature of the Phonological Code Accessed Early in Visual Word
Recognition during Sentence Reading
Brianna M Eiter, Ralph Radach, and Albrecht W Inhoff
ralph@psych.rwth-aachen.deState University of New York at Binghamton
There is a great deal of evidence which suggests that readers determine the phonological
repre-sentation of a word early during visual word recognition (Pollatsek, Lesch, Morris, & Rayner, 1992),
and that a word’s phonological code is maintained in working memory during sentence reading (Folk &
Morris, 1995) However it is unclear whether the phonological code used for a word’s identification is
also the code used in working memory storage The current study used a novel experimental method, the
contingent speech technique (Inhoff, Connine, & Radach, 2002), to examine this issue
Eye movements are used to trigger the presentation of a spoken word when the eyes reach a
par-ticular spatial location In Inhoff, Connine, Eiter, Radach & Heller (in press) readers spent more time
reading two words following the target when the spoken word was phonologically similar than when it
was dissimilar Presumably this occurred because the coactive phonological forms that developed on the
basis of the visual and auditory stimuli interfered with each other in phonological working memory thus
hampering post-target reading Interestingly, there was no similarity interference during the reading of
the visual target itself Articulation of the spoken word was relatively time consuming, and the prelexical
phonological code of the visual target had been determined before a functional segment of the spoken
word was articulated
The current study created experimental conditions in which a substantial segment of a
phonologi-cally similar and dissimilar companion word was articulated before a visual target was read An identical
condition, in which the articulated word and the target denoted the same word was used as baseline
Under these conditions, the similar spoken word was expected to interfere with both target and
post-target reading if a unitary phonological code was functional in post-target recognition and used later in the
target’s working memory storage In the experiment, the spoken companion word was presented either
10 character spaces before the eyes reached the target (the before condition) or immediately after the
target was read (after condition) The spoken word was identical, phonologically similar, or dissimilar to
the visual target In the before condition, target viewing durations (gaze durations) were shorter for
iden-tical auditory companion words in comparison to phonologically similar and dissimilar companions, and
the effects for similar and dissimilar companions did not differ The similar and dissimilar spoken word
interfered with target reading, relative to the after condition The identical condition yielded a small but
not reliable benefit in target word gaze duration Irrespective of when the spoken word was presented,
reading was impeded only for post target words in the phonologically similar condition This pattern
of results suggests that the phonological code used for visual word recognition is functionally specific,
i.e., that distinct types of phonological code are used for visual word recognition and for the word’s
representation in working memory after it has been identified
References
Folk, J R., & Morris, R K (1995) The use of multiple lexical codes in reading: Evidence from eye
movements, naming time, and oral reading Journal of Experimental Psychology: Learning, Memory,
and Cognition, 21, 1421-1429
Inhoff, A., Connine, C., Eiter, B., Radach, R., & Heller, D (in press) The phonological representation of
words after their fixation during sentence reading Psychonomic Bulletin & Review
Inhoff, A W., Connine, C & Radach, R (2002) A Contingent Speech Technique in Eye Movement
Re-search on Reading Behavior ReRe-search Methods Instruments and Computers, 34, 471-480
Pollatsek, A., Lesch, M., Morris, R., & Rayner, K (1992) Phonological codes are used in integratinginformation across saccades in word identification and reading Journal of Experimental Psychology:Human Perception and Performance, 18, 148-162
Trang 34Word processing, morphological surface structure and the lexicon
Catherine-Marie Longtin, Pierre A Hall´e, and Juan Segui
longtin@psycho.univ-paris5.frLaboratoire de Psychologie Exp´erimentale(CNRS, Universit´e Ren´e Descartes), France
There is growing evidence that morphological masked priming is sensitive to the morphological
surface structure of the prime and not to semantic transparency That is, morphological facilitation
ef-fects are obtained between two morphologically related words even if there is no semantic relationship
between them (Feldman et al., 2002; Longtin et al., 2003; Rastle & Davis, 2003; Rastle et al., 2000)
We will report a series of experiments using visual masked priming technique in which we investigated
how morphological structure and semantic transparency affect the processing of French words We used
pairs of words consisting of semantically transparent words and their base (fillette/fille ‘little girl/girl’),
semantically opaque words and their etymological base (vignette/vigne ‘label/vine’), pseudo-derived
words and their pseudo-base (baguette/bague ‘French bread, stick/ring’), and pairs of words that were
only orthographically related (abricot/abri ‘apricot/shelter’, -cot not being a suffix in French) Overall,
the results showed that, in masked priming (47 ms prime duration), a pure morphological priming effect
was obtained independently of semantic transparency, and not attributable to strict orthographic overlap
This pattern of priming was obtained when the surface morphological structure was controlled across the
transparent, opaque and pseudo-derived conditions (as in the examples above) and was replicated in two
different experiments
To explain these results, we will propose that morphological decomposition is a default analytic
process that is blind to the real morphemic status of the word and that applies to words having a surface
structure that is parsable into morphemes We propose that this ‘default decomposition’ applies at the
early stages of processing, which explains why it only surfaces in masked priming with very short prime
duration but not in a cross-modal paradigm (cf Longtin et al., 2003; Marslen-Wilson et al., 1994)
References
Feldman, L.B., D Barac-Cikoja, and A Kostic (2002) Semantic aspects of morphological processing:
Transparency effects in Serbian Memory and Cognition, 30(4), 629-636
Longtin, C.-M., Segui, J., Hall´e, P.A (2003) Morphological priming without morphological relationship
Language and Cognitive Processes, 18(3), 313-334
Marslen-Wilson, W.D., Tyler, L K., Waksler, R., & Older, L (1994) Morphology and meaning in the
English mental lexicon Psychological Review, 101 (1), 3-33
Rastle, K and M.H Davis (2003) Priming morphologically-complex words: Some thoughts from masked
priming Kinoshita, S and Lupker, S (Eds) Masked priming: The State of the Art New York:
Psy-chology Press
Rastle, K., Davis, M.H., Marslen-Wilson, W.D., & Tyler, L.K (2000) Morphological and semantic effects
in visual word recognition: A time-course study Language and Cognitive Processes, 15 (4/5),
507-537
Implicit Priming with Direct Word Naming in Single Word Production
Ching-Huei Tsai, and Train-Min Chen, Jenn-Yeu Chensun8514218@yahoo.com.tw, TrainMin.Chen@mpi.nl, psyjyc@ccu.edu.tw
National Chung-Cheng University, TAIWAN
The implicit priming task has served as an important paradigm in the recent studies of speech duction The task involves associative naming, whereby the participants first learn pairs of words that areassociatively or semantically related; afterwards, they have to name the target word when prompted withthe cue word in each pair The beauty of this paradigm is that the participants name the same words intwo different contexts, the critical manipulation of the paradigm In the homogeneous context, the words
pro-to be named share some morphophonological property (e.g., the first syllable) In the heterogeneous text, the same words are rearranged so that they no longer share the specified property When the namingtimes of the homogeneous context are subtracted from those of the heterogeneous context, the differencereflects only the effect of the context manipulation This is because the items serve as their own control,just like the participants do However, the paradigm has often been questioned for its possible involve-ment of other processes (e.g., memory) than word production Moreover, the task requires relatively longtime to complete In the present study, we investigated whether the function and the design beauty of theimplicit priming paradigm can be retained when associative naming is replaced by direct word naming
con-In Exp 1 and 2, the syllable and the tone of the first character of the target words served as the implicitprime The task variable (implicit priming with associative naming vs implicit priming with direct wordnaming) was manipulated as the between-subjects variable in Exp 1 and as the within-subjects variable
in Exp 2 In Exp 3, the first character/morpheme of the target words served as the implicit prime Taskwas a between-subject variable Results from all three experiments showed that although overall responsetimes were faster in direct word naming than in associative naming, the implicit priming effect was simi-lar in the two tasks We conclude that the implicit priming task with direct word naming taps on the samelevels of word production processes (specifically, word form encoding) as the implicit priming task withassociative naming The word naming task is easier and takes much less time to perform Therefore, itcan serve as a useful alternative for researchers interested in studying the word form encoding part ofspeech production
Trang 35The Computational Cost of Syntactic and Semantic Integration for
Processing Sentences with Relative Clauses — A Case Study of Chinese
Chin-Lung Yang1and Peter C Gordon2
cyang@pitt.edu
1 University of Pittsburgh, 2 University of North Carolina at Chapel Hill
The current study explores the universal as well as language-specific characteristics of the
pro-cessing and representation in memory of syntactic and semantic information over the course of language
comprehension by examining the relative ease of the processing of object-extracted and subject-extracted
RCs in Chinese Object-extracted RCs have been know to be harder to process than subject-extracted RCs
even though these two types of RCs have exactly the same linguistic constituents but differ only in word
orders as illustrated in the Example (1a) and (1b) below (Gibson, 1998; King & Just, 1991; MacWhinney
& Pleh, 1988)
1a Subject RC: [[The lawyer(i) that [e(i) attacked the politician]] stole the ballots]
1b Object RC: [[The lawyer(i) that [the politician attacked e(i)]] stole the ballots]
In contrast, in Chinese, the two types of construction are different from those of English in terms
of the ordering of the linguistic constituents, as shown in examples (2a) and (2b) (Note: The example is
shown with the English translation for each Chinese word DE is the marker for RCs in Chinese.)
2a Subject RC: [[[e(i) Attack(ed) the politician] DE the lawyer(i)] stole the ballots.]
2b Object RC : [[[The politician attacked e(i)] DE the lawyer(i)] stole the ballots.]
The crucial contrast is that while in English the RC comes after the head that is being modified;
in Chinese the RC comes before the head of the relative (e.g., in (1a) and (1b) the head noun “lawyer”
precedes the RC with attacked as the verb, while in (2a) and (2b) “lawyer” is modified by the RC but it
comes after the RC) Thus, the two languages present language comprehension processes with different
moment-to-moment tasks, and in turn offer a unique opportunity to test current theories of the cognitive
and linguistic processes that contribute to the understanding of the processing of complex sentences
The current study addresses this issue by conducting a series of self-paced reading-time experiments in
Chinese to examine the relative difficulty in Chinese of object-extracted and subject-extracted sentences
both when the RC modifies the subject NP of the matrix clause and when it modifies an object NP of the
matrix clause
The results suggest language-specific properties for the underlying mechanism of sentence
process-ing in Chinese While processprocess-ing Chinese RCs, the integration cost of the unintegrated sentence fragment
interacts with the computational cost of structural garden-path tapped by the incremental processing of
sentence comprehension in Chinese This pattern of results poses challenges for the cross-linguistic
gen-erality of current explanations of the differing processing difficulty between object- and subject-extracted
RCs that are based on primarily the computational cost and memory load imposed by different sentence
structures during comprehension (Gibson, 1998) The implication of the results will be discussed in light
of the evaluation of the relative merit and generality of contrasting theories in terms of object-subject
processing differences (Gibson, 1998; Gordon, Hendrick, & Johnson, 2001; King & Just, 1991; Lewis,
1996; MacWhinney & Pleh, 1988)
References
Gibson, E (1998) Linguistic complexity: Locality of syntactic dependencies Cognition, 68, 1-76.Gordon, P., Hendrick, R., & Johnson, M (2001) Memory interference during language processing Journal
of Experimental Psychology: Learning, Memory, and Cognition, 27(6), 1411-1423
King, J & Just, M A (1991) Individual differences in syntactic processing: The role of working memory.Journal of Memory & Language, 30(5), 580-602
Lewis, R L (1996) Inference in short-term memory: The magical number two (or three) in sentenceprocessing Journal of Psycholinguistic Research, 25, 93-115
MacWhinney, B., & Pleh, C (1988) The processing of restricted relative clauses in Hungarian, Cognition,
29, 95-141
Trang 36Investigating Gap-Filler Dependencies in Chinese: Is There an ‘Active
Gap’?
Chun-chieh Natalie Hsu and Benjamin Brueningcchsu@udel.edu, bruening@udel.eduUniversity of DelawareStudies on the processing of English have shown that when the parser encounters a filler, it actively
looks for a gap (Crain and Fodor 1985, Stowe 1986, etc.) This “active filler” effect is evidenced by the
reading time at the potential gap position filled by ‘us’ in (1b) being longer than the reading time to ‘us’
in (1a), where no such dependency exists
(1) a My brother wanted to know if Ruth will bring us home to Mom at Christmas
b My brother wanted to know WHO Ruth will bring us home to GAP at Christmas
Few studies, however, have examined constructions where the gap precedes the filler This study
inves-tigates gap-filler dependencies in Mandarin Chinese prenominal relative clauses, where the head noun
follows the relative clause that modifies it, as in (2b) We use a filled-gap paradigm like that in (1), with
an optional adjunct phrase occupying the first potential position for the filler If the parser actively seeks
a filler for the relative clause gap in (2b), the reading time across the adjunct phrase should be longer in
(2b) than in (2a), where there is no gap, since that position is a potential spot for a filler
(2) a Na-wei lao-taitai zuotian bianzhi-le yi-jian maoyi
that-CL old-lady yesterday knit-PAST one-CL sweater
‘That old lady knitted a sweater yesterday.’
b Na-wei [GAP zuotian bianzhi-le yi-jian maoyi ] de LAO-TAITAI shebing-le
that-CL yesterday knit-PAST one-CL sweater DE old-lady get-sick-PAST
‘That old lady who knitted a sweater yesterday got sick.’
c Na-wei lao-taitai zuotian bianzhi-le yi-jian maoyi suogei ta-erzi
that-CL old-lady yesterday knit-PAST one-CL sweater give-to her-son
‘That old lady knitted a sweater to give to her son yesterday.’
d Na-wei [GAP zuotian bianzhi-le yi-jian maoyi suogei ta-erzi ] de LAO-TAITAI shebing-le
that-CL yesterday knit-PAST one-CL sweater give-to her-son DE old-lady get-sick-PAST
‘That old lady who knitted a sweater to give to her son yesterday got sick.’
Analyses of reading times reveal that each word in the adjunct phrase as well as the adjunct phrase as a
whole are read significantly more slowly in (2b) than in (2a) These results support the ‘Active Gap’
hy-pothesis, which says that the parser must actively seek a filler to satisfy a gap held in memory They also
demonstrate processing effects of long-distance filler-gap (or gap-filler) dependencies in Chinese;
more-over, effect in Chinese gap-filler constructions is the same as the effect in English filler-gap constructions,
showing that the strategies used by the human language parser in processing long-distance dependencies
are the same regardless of language or directionality Finally, the finding that filler-gap/gap-filler
de-pendencies are processed uniformly supports memory- or prediction-based theories of processing (e.g.,
Gibson 1998)
References
Crain, S., and Fodor, J.D (1985), ‘How can grammars help parsers?’ In D.Dowty, L Karttunen, and A
Zwicky, eds., Natural language parsing: psychological, computational, and theoretical perspectives,
Cambridge University Press, pp 94-128
Gibson, Edward (1998), ‘Linguistic Complexity: Locality of Syntactic Dependencies.’ Cognition 68: 1-76
Stowe, L (1986), ‘Evidence for on-line gap location.’ Language and Cognitive Processes 1: 227-245
Contextual effects on the time course of linguistic and conceptual gender information in understanding double-gender words
Cristina Cacciari1, Roberto Padovani1, Alberto Verzellesi1& Manuel Carreiras2
cacciari.cristina@unimore.it, padovani.roberto@unimore.it, mcarreir@ull.es
1 University of Modena, Italy; 2 University of Tenerife, Spain
Epicene words (e.g., “vittimaFEM”, “victim”) are Italian words syntactically marked as either inine or Masculine At a discourse level, they can be used to refer to both male and female individuals.What makes epicenes interesting is that they allow to distinguish between a linguistic level of genderrepresentation and a conceptual one In Cacciari et al (1997), epineces were investigated when theyacted as antecedent for diverse types of anaphor embedded in contexts with no bias as to the gender ofthe referent Participants were always faster in resolving an anaphor whose gender matched the syntac-tic gender of the epicene In Cacciari and Carreiras (2001), a similar paradigm was employed but theepicenes were embedded in contexts that provided information on the gender of the referent When thecontext biased toward a referent’s gender syntactically consistent with the epicene, a reading time ad-vantage emerged for the pronoun matching both the conceptual and the linguistic gender of the epicene.When the contextual bias was reversed, faster reading times were obtained with pronouns conceptuallymatching the gender of the intended referent with respect to pronouns linguistically matching the gender
Fem-of the epicene These results suggest that linguistic and conceptual gender represent two different sources
of information differently used in anaphor resolution
The aim of the two experiments presented here is to explore the role of linguistic and conceptualinformation in assigning an antecedent to a pronoun when it corefers either with an epicene or with abigender word that acts as a control condition having no explicit gender mark associated (e.g., “erede”,
“heir”) We used a non cumulative self-paced moving window for measuring reading for comprehensiontimes In Experiment 1, neutral, gender-congruent and gender-incongruent contexts were used that endedwith a sentence whose referent in the subject position was either an epicene or a bigender word Thepronoun of the critical sentence was always syntactically congruent with the epicene In Experiment
2, the same context were used but the critical segment was varied in that a pronoun mismatching thesyntactic gender of the epicene, and matching the conceptual gender biased by the context, was used
References
Cacciari, C., Carreiras, M & Barbolini Cionini, C (1997) When words have two genders: Anaphor lution for Italian functionally ambiguous words Journal of Memory and Language, 37, 517-532.Cacciari, C & Carreiras, M (2001) The role of linguistic and conceptual gender in Italian pronoun reso-lution Presented at the Annual CUNY Conference, Philadelphia, USA
Trang 37reso-Overt prosody in the RC-attachment construction: Elicitation protocols
Dianne Bradley1, Eva M Fern´andez2, Nenad Lovric1
dbradley@gc.cuny.edu, eva fernandez@qc.edu, nlovric@hotmail.com
1 Graduate Center, CUNY; 2 Queens College & Graduate Center, CUNY
The “Post-to-Times” protocol of Bradley, Fern´andez & Taylor (2003) (BFT) presents two short
sentences, see (1), to elicit an utterance containing a complex NP with a modifying relative clause
(RC); for the speaker, RC’s attachment is disambiguated BFT reported that their instrumental analyses
of elicited utterances (N2-disambiguated, uniformly) showed remarkable systematicity in phonological
phrasing: Whole-sentence length controlled the likelihood of phrasal breaks occurring at RC’s left edge,
i.e., N2][RC They argued that the overt prosody facts support an implicit prosody explanation (Fodor,
2002) of RC-attachment preferences: When ambiguous sentences were read silently, attachment was
higher both when matrix subjects were heavier and when RCs were longer; see (2)
We report research extending these preliminary findings We first examine BFT’s claim that N2][RC
is the sole site of systematic variation in default phonological phrasing in English because it is privileged
in the syntax/prosody interface of that language We evaluate the possible objection that this break site
has merely been picked out by a protocol presenting N2 sentence-finally Data were collected in an overt
prosody study of Croatian, a language in which a proclitic preposition ’od’ (non-thematic, and similar
to English ’of’) optionally precedes N2 in the complex NP; Lovric (forthcoming) shows N1][Prep-N2
to be a second site attracting phrase breaks in Croatian’s default prosody With materials factorially
combining RC-Length (short/long) and Preposition (absent/present) and utterances elicited with BFT’s
protocol (see (3)), we demonstrate that phonological phrasing for this construction in Croatian involves a
trade-off between two break-sites: RC’s left edge, as in English, and Prep-N2’s left edge BFT’s findings
for English are not, therefore, an artefact of the protocol
In a second study, we explore effects of RC’s (non-)restrictiveness on phonological phrasing in
English, contrasting two variants of the elicitation protocol Restrictive RCs are elicited when an
intro-ductory sentence is accompanied by a “Which X?” question and response, as in (4a) and (4b), and
non-restrictive RCs, when (4c) accompanies (4a); note that RC predicates are segmentally identical across
restrictive and non-restrictive types We demonstrate that it is only for restrictive RCs that the likelihood
of the N2][RC phrasing break reliably grades with whole-sentence length This result suggests an
im-plicit prosody account of the finding of Hemforth et al (submitted), that extraposed RCs in German fail
to exhibit length effects on preferred attachment Separate phonological phrasing of RC is obligatory
under extraposition; and where break-likelihood is at ceiling, length-sensitivity is ruled out
(1) The plot concerns the guardian of the prince The (guardian/prince) was exiled
(2) The (unusual) plot concerns the guardian of the prince who was exiled (from the country for
decades)
(3) Opisali smo bratica (od) rukometasa
Described are-[1st,PL] the cousin (of) the handball-player
Rukometas studira (na odsjeku za arheologiju)
The handball-player studies (at the department of archeology)
(4) a The plot concerns the guardian of the prince
b Which prince? The prince who was exiled
c By the way, that particular prince was exiled
References
Bradley, D., Fern´andez, E & Taylor, D (2003) Prosodic weight versus information load in the relativeclause attachment ambiguity Paper presented at the 16th Annual CUNY Conference on HumanSentence Processing, Cambridge MA
Fodor, J D (2002) Psycholinguistics cannot escape prosody Proceedings of the 1st International ence on Speech Prosody, Universit´e de Provence, 83-88
Confer-Hemforth, B., Fern´andez, S., Clifton, C., Jr., Frazier, L., Konieczny, L., & Walter, M (submitted) Relativeclause attachment in German, English and Spanish: Effects of position and length
Lovric, N (forthcoming) Implicit prosody in silent reading: Relative clause attachment in Croatian toral dissertation, CUNY Graduate Center
Trang 38Doc-Verb activation patterns in Dutch matrix clauses during on-line spoken
It has been proposed that in non-canonical structures a noun that is displaced from its canonical
position is reactivated on-line at that gap position (e.g Love & Swinney, 1996; Nagel, Shapiro, & Nawy,
1994), suggesting that listeners must connect non-adjacent positions on-line in order to interpret such
constructions While linguistic theory proposes both noun and verb movement, it is unknown whether
displaced verbs are recovered (re-activated) in their canonical position in the same manner as are NPs
We present two cross-modal lexical priming experiments exploring this issue In Dutch declarative
matrix clauses the finite verb has moved from its canonical clause final position to the second position,
leaving behind, theoretically, a gap Therefore, Dutch matrix clauses provide a good opportunity to study
verb activation The experimental sentences in our experiments consisted of a matrix clause (SVO)
fol-lowed by an embedded clause (see examples below) Our lexical decision probes were verbs related to
the moved verb, unrelated matched controls and non-words derived from verbs We predicted faster RTs
to verb-related relative to unrelated control probes both directly after the verb and at the gap, if verbs are
indeed (re)activated at the gap as has been found for nouns
In the first experiment we tested for activation of the verb at three probe-points: [1] directly after
the verb, [2] 700 milliseconds downstream from the verb, and [3] directly after the complementizer
(where the parser unambiguously knows that there was a gap) To further examine the time-course of
verb activation, in the second experiment we tested at four probe-points: [1] directly after the verb, [2]
1500 ms downstream from the verb, [3] directly after the object head noun (where the gap associated
with the moved verb is located) and [4] after the complementizer
The results suggest that verbs behave differently from nouns We found no evidence for reactivation
of the verb at its canonical position Rather, the verb remained activated from the point where it was first
encountered until the end of the clause (significant priming effects were found directly after the verb, at
700 ms and 1500 ms downstream from the verb as well as at the offset of the object head noun, while no
priming effects were found after the complementizer) We propose that the verb is kept active to ‘find’
its arguments in order to theta-mark them Once thematic roles are discharged, we assume that activation
of the verb dissipates
Experiment 1
De wanhopige verslaafden beroven[1] eenzaam wande[2]lende bejaarden, omdat[3]
The desperate addicts rob[1] lonely stro[2]lling seniors, because[3]
Love, T & Swinney, D (1996) Coreference processing and levels of analysis in object-relative
construc-tions: demonstration of antecedent reactivation with the cross-model priming paradigm, Journal of
Psycholinguistic Research, 25 (1), 5-24
Nagel, H.N., Shapiro, L.P., & Nawy, R (1994) Prosody and the processing of filler-gap sentences Journal
of Psycholinguistic Research, 23, 473-485
Frequency can account for word order effects in German: A matter of
the appropriate corpus query
Dirk P Janssen, Sandra Muckel, and Johannes Schliesserdirkj@uni-leipzig.de, muckels@uni-leipzig.de, schliess@uni-leipzig.de
Institut f¨ur Linguistik, Universit¨at LeipzigWord order effects in sentence processing can be interpreted as structural preferences of the parser
or as a result of the amount of experience with syntactic structures Some recent studies have presenteddata of which the authors claimed that they could not be the result of frequency of occurrence and aretherefore parser properties (Bornkessel, Schlesewsky and Friederici, 2002; Scheepers, Hemforth andKonieczny, 1999; R ¨osler et al, 1998)
We decided to look into this issue further by running an exhaustive query on the NEGRA Germannewspaper corpus (http below) We wrote a corpus parsing tool that located the arguments of the mainverb, including antecedents, and removed all non-argument objects from the tree Crucially, our tool isable to retrieve arguments of the main verb from nested parts of the tree, as in
(S (NP-nom Sandra)(V will)(ADV maybe)(VP-clause (V give) (NP-dat the Dutchman) (NP-acc the book)))
which will be counted as a nominative–dative–accusative structure
In contrast to normal matching or cooccurrence methods, our program analysed the argument ture of each sentence in full We claim this method gives much more precise results, because it is notsensitive to intervening sentence elements, it can handle arbitrarily complex structures including con-junctions, it can distinguish pronominal elements from full NPs and it will try several alternative parses
struc-of the same sentence to find the optimal parse
Contrary to what Scheepers et al claimed, we found much larger incidence of structures with twoarguments than with three, both when all arguments were full NPs and when pronouns were included.The interpretation of the data presented by Bornkessel et al is more complicated: Depending on whetherone excludes sentences with non-final verbs or sentences with pronouns, we find numbers that contradict
or confirm Bornkessel et al.’s counts The theoretical question here is whether frequency of occurrenceneeds to be defined on the basis of the precise sentence make-up used in the experiment, or whether oneneeds to be more permissive
We conclude that when making claims about parser-preferences, one needs to use reliable corpusfrequencies to exclude the possibility that a occurrence based explanation can be made More specifictheories should be developed about whether frequency of occurrence is defined on the basis of argumentorder per se, or on argument order in specific lexical-syntactic contexts
lan-Scheepers, C., Hemforth, B., & Konieczny, L (1999) Incremental processing of verb-final constructions:Predicting the verb’s minimum (!) valency Paper presented at the International Conference on Cog-nitive Science, Tokyo
Trang 39Computer Simulation of a Serial Parsing Model
Doan-Nguyen Haihaidoan@cs.concordia.caConcordia University
We implemented a computer simulation of a serial model of syntactic parsing, based on the
garden-path theory (Frazier, 1978; Frazier & Clifton, 1996) The parsing process is basically incremental and
input-driven (bottom-up) Attachment of an incoming word is first (Step 1) guided by whether there is
a construction in the current partial parse tree (CPPT) needing or expecting a constituent (eg a verb
waiting for an object, or a postulated CP waiting for its TP complement) Next (Step 2), it is guided by
specific properties of the incoming word and CPPT, and preference principles like Minimal Attachment
and Late Closure If attachment fails, reanalysis (Step 3) is invoked Based on Fodor & Inoues (1994)
Diagnosis model, reanalysis is implemented following the formula:
Reanalysis = Symptoms + Guesses + Verification + Repair
In fact, reanalysis is not necessarily invoked only when attachment fails This happens when
at-tachment is still possible but less probable than a reanalysis option We found at least two such cases:
multiple subcategorization verbs and object traces Eg.,
1 I want him to come (Not to attach to come to want as a purposive adjunct, but reanalyze [VP
want [DP him]] into [VP want [CP [TP [DP him] to [VP come]]]].)
2 The pictures(i) you have e(i) seen (Not to attach seen as a modifier of the pictures (cf The
pictures you took seen from this perspective are very surrealist), but combine it with have to make a
perfect tense, pushing the trace e(i) to the object position of ’seen’.)
These cases show a need of adding a step between Step 1 and Step 2 of the algorithm (say Step 1.5),
which checks the possibility of doing a reanalysis rather than an ”easy” attachment To help speed up the
check, the parser memorizes, for the first case, the other potential subcategorizations of a verb whenever
it is satisfied with a subcategorizational configuration (cf Stevensons (1998) decay thesis), and, for the
second case, the position of a trace whenever it is located When a trace is pushed out of its current
position by reanalysis, it can be moved into a new position, as in:
3 The pictures(i) I want e(i) to see (e(i) moves to the object position of see)
or hung out waiting for a new position, as in:
4 The folder(i) I kept e(i) my letter in is blue
(e(i) is hung out when my letter pushes it out of the object position of kept, and finally located
when in arrives.)
The parser has some top-down mechanisms, conform to a general agreement in recent
psycholin-guistic research that parsing is not totally bottom-up but has also top-down features (Frazier, 1998)
Besides the well-known structure postulation mechanism, we propose forward parsing, in which the
parser delays attachment and continues to work with the input stream to gather enough information for a
good attachment Eg., forward parsing is used to find the head of a DP to avoid repeated reanalysis as in
the case of:
5 We need storage management software
6 I saw the teacher’s son’s dog
The current prototype covers many basic and advanced structures of English, including noun, verb,adjective, and preposition phrases, adjunct structures, relatives, yes-no and wh-questions, and tracechains Tests done on a set of 300 sentences of length up to 25 words (at least 10 sentences for eachlength) yielded a linear execution time with respect to input length, with much longer time for reanaly-sis cases than for straightforward cases This shows that the implementation apparently conforms to itsobjective to simulate a human parsing model
References
Fodor, J.D & Ferreira, F., 1998 (eds) Reanalysis in Sentence Processing Kluwer Academic
Fodor, J.D & Inoue, A., 1994 The Diagnosis and Cure of Garden Paths Journal of PsycholinguisticResearch, Vol 23, No 5, 1994
Frazier, L., 1978 On Comprehending Sentences: Syntactic Parsing Strategies Doctoral dissertation, versity of Connecticut
Uni-Frazier, L., 1998 Getting There (Slowly) Journal of Psycholinguistic Research, Vol 27, No 2, 1998.Frazier, L & Clifton, C., 1996 Construal Cambridge, MA: MIT Press
Stevenson S., 1998 Parsing as Incremental Restructuring In Fodor & Ferreira (1998) 327-363
Trang 40Why ‘that’?
Douglas W Roland; Jeffrey L Elman; Victor S Ferreiradroland@crl.ucsd.edu, elman@crl.ucsd.edu, ferreira@psy.ucsd.edu
University of California, San Diego
The direct object / sentential complement ambiguity which occurs at the post-verbal NP in Tom
admitted the students were right can be eliminated by including the complementizer that Yet production
experiments such as Ferreira and Dell (2000) suggest that speakers do not use that to avoid such
am-biguities, and may not even be aware of the ambiguity during production At the same time, evidence
suggests that there are systematic differences between examples that do and do not use the
complemen-tizer Thompson and Mulac (1991) show that the epistemicity of the main clause and the topicality of
the complement contribute to that ellipsis Hawkins (2002) argues that ellipsis is less likely when the
subordinate subject is long, due to processing pressures to avoid lengthy ambiguity, and Ferreira and
Dell (2000) suggest that ellipsis occurs to allow early mention of available material, such as when the
main and subordinate clause subjects are the same
These previous analyses have involved relatively small data sets, large enough to support the idea
that each factor is a potential predictor of that ellipsis, but too small to allow for detailed examinations of
the relative strength of each cause We prepared a database of the approximately 1.3 million spoken and
written tokens from the British National Corpus containing any of the 100 sentential complement-taking
verbs listed in Garnsey et al (1997) Approximately 182,000 of these instances involved a sentential
complement, either with or without that These examples were labeled for a variety of formal and
se-mantic properties The formal properties included the length of the subject and post verbal NPs and their
heads, and the log lexical frequency of the heads of the subject NP and the post-verbal NP The semantic
properties consisted of automatically ranking the subject and post-verbal NPs and their heads on twenty
semantic dimensions based on Latent Semantic Analysis (Deerwester et al 1990) We then performed
a variety of analyses on this data, including regression analyses to predict the extent to which various
factors could predict the presence or absence of that
Starting from a baseline of 63%, we were able to correctly predict the presence or absence of a that
79% of the time This provides clear evidence that the presence of the complementizer is governed by
contextual factors Additionally, we show these factors conspire to reduce the occurrence of ambiguous
cases such as the example in the first sentence
References
Deerwester, S., Dumais, S T., Furnas, G W., Landauer, T K., & Harshman, R (1990) Indexing by Latent
Semantic Analysis Journal of the American Society For Information Science, 41, 391-407
Ferreira, V S., & Dell, G S (2000) Effect of ambiguity and lexical availability on syntactic and lexical
production Cognitive Psychology, 40(4), 296-340
Garnsey, S M., Pearlmutter, N J., Myers, E., & Lotocky, M A (1997) The contributions of verb bias
and plausibility to the comprehension of temporarily ambiguous sentences Journal of Memory &
Language, 37(1), 58-93
Hawkins, J A (2002) Symmetries and asymmetries: their grammar, typology and parsing Theoretical
Linguistics, 28(2), 95-150
Thompson, S A., & Mulac, A (1991) The discourse conditions for the use of the complementizer ‘that’
in conversational English Journal of Pragmatics, 15, 237-251
Quantifying word order freedom in natural language: Implications for
sentence comprehension
Geert-Jan M Kruijff and Shravan Vasishthgj@coli.uni-sb.de, vasishth@coli.uni-sb.deUniversity of Saarland
A robust, unsupervised approach is presented for learning word order models from treebanks forlanguages with variable degrees of word order freedom The approach is based on an extension of profileHidden Markov Models (pHMMs; Durbin et al 1998) A pHMM builds a model that represents similar-ities across the structures it is trained over; when trained using syntactic trees, we obtain a model of whatconstituents may appear in different (relative) positions This represents a generalization of topologicalfield models (Hoehle 1983), one with which we can model restrictions on and possible variations in wordorder
This approach is evaluated using treebanks for English (Wall Street Journal), Dutch (Corpus of ken Dutch), German (NEGRA), and Czech (Prague Dependency Treebank) We show that it is possible
Spo-to define typological differences in word order freedom along a continuous scale of entropy distanceintervals (EDI); this is in contrast to the binary or ternary categories of traditional typological research(e.g., Steele 1978)
Such a continuous scale for defining degree of word order freedom has interesting consequencesfor real-time language comprehension The more rigid the word order of a given language along the EDIscale (e.g., English), the more susceptible it should be to increased processing difficulty when noncanon-ical orders are encountered Conversely, the processing of languages located towards the freer end of theEDI continuum (e.g., Czech) should be affected to a lesser extent (or not at all) by noncanonical orders
In order to investigate this prediction, acceptability-rating and self-paced reading experiments volving German and Czech are currently in progress that investigate the effect of noncanonical wordorder on processing difficulty The EDI scale predicts that noncanonical order in German would be lessacceptable (even with appropriate discourse context present) than noncanonical order in Czech
in-In sum, a metric is proposed for quantifying word order freedom across languages, and it is arguedthis metric can correctly predict the differentiated degrees of dispreference for noncanonical order inlanguages conventionally classified as “free word order” languages
References
Richard Durbin, Sean Eddy, Anders Krogh, and Graeme Mitchison 1998 Biological Sequence Analysis.Cambridge University Press
Tilmann Hoehle 1983 Topologische Felder PhD thesis University of Cologne
Geert-Jan M Kruijff A Categorial-Modal Logical Architecture of Informativity: Dependency GrammarLogic and Information Structure 2001 PhD thesis, Charles University, Prague, Czech Republic.Susan Steele 1978 Word order variation: A typological study In Joseph H Greenberg, editor, Universals
of Language, Volume 4: Syntax, pages 585-624 Stanford University Press, Stanford, CA