Báo cáo khoa học: "EXTENDED ACCESS TO THE LEFT CONTEXT IN AN ATN PARSER" pptx

EXTENDED ACCESS TO TRE LEFT CONTEXT IN AN ATN PARSER Irina Prodanof and Giacomo Ferrari Istituto di Linguistica Computazionale Via della Fageiola 32 1-56100 Pisa, Italy ABSTRACT Some It

EXTENDED ACCESS TO TRE LEFT CONTEXT IN AN ATN PARSER

Irina Prodanof and Giacomo Ferrari Istituto di Linguistica Computazionale

Via della Fageiola 32 1-56100 Pisa, Italy

ABSTRACT

Some Italian sentences related to linguistic

phenomena largely known and recently discussed by

many computational linguists are discussed in the

framework of ATX They offer certain difficulties

which seem to suggest a substantial revision of

the ATN formalism The theoretical assumptions and

an experimental implementation of such a revision

are presented, together with examples Hany

related theoretical points such as sone

psycholinguistic implications and the relationship

between deterministic and non-Jeterministic

hypothesis are also briefly discussed

I INTROOUCTION

Certain types of sentences seem to defy the

abilities of several parsers, and sowe of them are

beinz now discussed by wany computational

linzuists, mostly within the deterministic

hypothesis

An examination of their treatment within the

traditional ATE paradigm seems to suggest that the

real discussion is about how to acces the left

context and what its form should be

II ACCESS TO THE LEFT CONTEXT

A <ATN Grammars

An ATN grammar is a set of networks formed by

labelled states and directed arcs connecting then

The arcs can ricopnize terminal (words) and

non-terminal (lexical caterories) symbols or

recursively call for a network identified by the

jdabel of an initial state When such a call

(1) The ambixuity of this sentence is the sane as

occurs, the parsing algorithn suspends the current computation and starts a new level of computation Usually, each network recognizes some linguistic unit such as Noun Phrase (NP), Prepositional Phrase (PP), and Sentence (8S) and any recursively embedded call to one of them corresponds to a level of computation

The parsed parts of the input stored (SETRed) into resisters as recopnized AÁt the end of tlhe network registers are combined (BUILDGed) into a parse node and returned (POPed) to the calling net Appropriate functions can return the content of a register (CETR) or transfer it to another register (combination of SETR with a GETR) This last operation is equivalent to i) the renaming of a register, if the source register is successively set to a different value, ii) the initialization

of a register at a lower or higher level, if SENDR

or LIFTR are used

string they are are these

Initialization is commonly used to i} raise lexical features to a higher level where they are used for tests (ex.: subject-verb agreement), ii) pass possible antecedents to lower levels where a gap may be detected in an embedded clause

Be Difficult access to registers

1 Filler-gap Linking The antecedent passing theoretically unlimited increase

By the standard procedure, the ambiguous sentence(1)

may cause a

in storage load analysis of the

(1) Giovanni disse che John said that (he) had

aveva mentito lied

“Giovanni” is always SENDRed as possible SUBJect

of a complement, as soon as “disse” is recognized

as an STRANS verb As no subject NP is wet after

“che”, an interpretation is yielded with

“Giovanni” in subject position The second interpretation is produced sinply by successively setting the SURJ register to a duamy node, which

its English translation where “he* can be bound either

tơ “John® or to souweone else mentioned in a previous sentence Italian has a gap instead of a pronoun

renains unfilled The sane treatment is

recursively applied to sentences Like

(2) Giovanni pensava che avrebbe raccontato

John thought that (he) would have told

a tutti che - aveva fatto una

to everybody that (he) had done a

scoperta

discovery

where “Giovanni” must serve as subject of both the

first and the second (linearly) complement

Instead, in the sentence

(3) Giovanni disse che i suoi colleghi avevano

John said that his colleagues had

mentito

lied

as the NP “i suoi collephiF is analysed, it

replaces the SENDRed “Giovanni” in the SUBJ

register and the correct interpretation is popped

A ere

sentence

couwplex Creatuent is required for the

(4) Giovanni pensava che i suoi collerhi avrebbero

John thought that his colleagues would

raccontato a tutti che aveva fatto

ave told to everybody that (he) had done

una scoperta

a discovery

where “Giovanni* must get through the first

complement and reappear in the second (embedded)

one If SENDR is used, a transit repister Ri with

the same content as the initialized SUBJ register

is to be passed down torether with it When the

subject of the first complement (“i suol

collerghi”) is found, it replaces “Giovanni” in

SURJ but not in Rl The new SUBJ “i suoi

colleghi” and Rl “Giovanni” are again SENDRed to

the embedded complement where the agreement with

the verb correctly selects, as subject, the

content of Kl ow, as the number of levels which

to “juập over” is in principle illinited and each

one way have its own subject to be SENDRed, a

transit repister for each new subject is needed

Thus, for the seutence

(5) Giovanni era sicuro che i suoi nenmici

avrebpero rivelato alla staapa che sua mo;lie

aveva detto un siorno che l*aveva picehiata(2)

wher processing the laste couplement three

recisters containin,, the three possible subjects

Will be available an:l sliall be visited in order to

find the rirht one(3)

(2) “Jolin was sure that his ecnnenies would have disclosed

2 Lexical features raising

A storage overload way also be caused by the need for raising lexical features Morphological features are necessary at the least in the test of subject-verb agreement This is done by LIFTRing

in ad hoc registers gender and number from the NP level to the S level If the NP is popped ina possible subject position the test for agreewent may take place by comparing the content of those Tegisters with the corresponding features of the verb However, there are cases such as ex.(1) in which such information must be used again in the course of the analysis for another (agreement) test Those features must be, therefore, copled

in appropriately labelled repisters in order to 1) flag their relation to the subject and ii) prevent then from being erased when the same features are LIFTRed from the following NP

The same need for ad hoc storing may be shown for the object ana few PP complements For example, in the sentence

(6) Il capitano disse ai soldati che The captain said to the soldiers that marciavano pietosamente

(they) were marching peatifully

as the NP 7i soldati*® embedded in tiie “a-PP” is the subject of the complement, its vender and number must travel up through PP and S$ and down to the complement, in order the subject-verb apreement to be satisfied This implies that at a given moment other registers containing uorphological features coexist at the § level with those of the subject

An obvious remedy to this proliferation of repisters consists in adding these features to the

NP or PP structure, POPing them, and extracting them from those structures when needed Sut classical ATN does not provide a suitable function, a form that should return only a part

of a stored tree after having searched through it Now, although such functions are not particularly complex to desinn [5, 7], they are not perfectly consistent with the ATN general formalism which does not take into consideration the extraction of pieces of already processed information from the Structure they have veen inserted in

C <A functional perspective

1 A generalized retrieving action

If we look at the discussed exauples from an entirely functional viewpoint we can describe theu

as having in common the need for retrievin;

to the press that his wife had once told that (he) had beated her® ile ive here, for clarity, the parenthesized fora of this example:

(Giovanni era sicuro (che i

giorno (che 1“aveva picchiata)))

Notice that in this

this is not always the case

suol nemici avrebbero rivelato example the subject-verb afreement is sufficient to select

alla stampa (che sua moplie aveva detto un

the rislt antecedent, bue (3) A pessible alternative, equally tricky,is the use of the HOLD=VIRT couple

59

information somewhere back in the already built

structures; the tricky solutions presented above

are, in fact, a way of accessing parts of the left

context These sometimes correspond to the entire

content of a register and sometimes to a fragment

We will assume, then, that the left context is

stored in a space of memory, equally accessible

from any level and that retrieving always concerns

fragments of it At any point of the process this

Structure contains the current hypothesis about

the analysis of the parsed secment of the input

from the besinning; hence we will refer to it as

Current Global Hypothesis (CGH)

The retrieving action will have two participants,

a symbol that triggers the action (trigger) and

the information to he retrieved (the target of the

action)

In this frame all the different procedures

discussed above may be reduced to a sinsle seneral

alsorithm of three steps,

1) identification of a tricrer (a fap to be

filled, a verb which demands for the subject-verb

acreement test)

ii) extraction of constraints which must suide the

search for the tarpet, and

iii) retrieving of the required inforuation

the same

an anaphoric

On this functional ground,

description fits to the binding of

pronoun to its antecedent [7] as in

(7) a) Giovanni disse a ‘aria che voleva

Jón said to Mary that (he) wanted

sposarl-la

marry her

as opposed to

b) Giovanni disse a liaria che voleva

John Said to Nary that (he) wanted

Sposarsi

fo marry

The function that searches bach nay be

constrained hy several types of restrictions,

including i) morphological features, i.e the

gender and number of the pronoun or those required

for ayreement by the syntactic environment (e-.2

the verb), ii) syntactic idiosyncrasies of soue

lexical item such as the STRANS verbs that

determine which of their arguaents is to be the

Subject of the couplement, iii) setantic features

that way be introduced in the process, and iv)

syntactic deternination of the scope of the

search

2 ,laninularions o£ the left context

Ketrievin: of the antecedent may actually

correspond to two different

upon whether the antecedent to be

preceeds or follows the syubol it is

operations dcpendinys bound linearly

to be bound

to

In the cases discussed above the zap or the

pronoun always follow there antecedent and it is

therefore possible to imaediately insert the

binding pointer lioreover, in wany couwon sentences the antecedent linearly follows its dependent, as

in (3) Quando — sỉ arrabbia, Giovanni diventa rosso When (he) gets angry, John becomes red (9) Se lo vedi, saluta Giovanni

If (you) see him, say halio to John

da parte mia

on ny behalf

In this case, the bindings two steps, the flagging of

should take place in the need for a forward binding and the moving of the pointer from the antecedent, once detected, to the flag Notice that this class of sentences entirely lies beyond the abilities of the classical ATi: framework Relative pronouns also need to be bound to an antecedent and, besides, are tie surface signal of

an embedding No special processing difficulty is proposed by sentences like

(10) I] rarazzo che corre The boy who runs where the relative pronoun occurs emactiy where the embedding benins In this case ai scope restriction can limit the searcu for an antecedent

to the inmediately preceeaing kP

But in the sentences (11) Il ragazzo del quale ti parlavo The boy about whom I was talkin; to you (12) Il ragazzo del cui padre ti parlavo The boy about whose father I was talking

to you the relative clause boundery is set one word before the relative pronoun, and in

(13) 11 ragazzo del padre del quale The boy about the father of whom

ti parlavo

1 was talking to you where the phenowenon known as pied-pipins occurs, such a boundery may be located several words before the relative pronoun

In an ATH these cases cect the initial set of arcs which recojnize a PP ewbeJded in an NP, as in (14) il rayazzo dell*ascensore

the boy of the lift and the correct interpretation is obtained only by backtracking, (for Ll and 12) and asain using, a set

of transit resisters (for 12) No solution is known for sentence (13)

discussin;) an wodifies the left should ented = the

a relative clause

In the framework we have been action which structurally context can be proposed It component(s) being processed in

as the relative pronoun in net

A third type of access to the left context is

the relabelling of a processed component, already

used for the passive transformation

TTI 23⁄PENIIFEXTAL TIPLEIIENTATIOI

A General assumptions

ilost of the above discussion is based

own expericnce with an

also for other parsers

phenomena we are trying to

on our ATN parser, although valid Some of the linguistic functionally classify have been recently discussed in the frame of the

deterainistic parsing and sowe progress in the

treatment of those sentences has been done

However we prefer to stick to the

non-detemiinistic hypothesis, because no valuable

argument seems to stand against the idea of ATN as

a collection of alternative heuristic strateyies

representing those usel in the human = sentence

processing Un the contrary it is possible to

introduce in an ATN parser soue nechanisn for

selectins stratesies in function of a text ora

subBlansuaze [11]

Our assumption is, then, that the possibility

of backtrackins must be kept, but restricted by

1 introducing linpuistic constraints to ,;ulde the

choice of alternatives;

iil manipulating the left context in some cases in

which backtracking can be intuitively charged on

inadequacies of the parser rather than on points

of real linjuistic ambiguity

A second asswaption wore related to the

current inplenentation of the systeu refers to the

pranunatical formalisn A functional representation

a la ¡say {14] is used A detailed explanation

of this fonaalisa is not necessary to the

understandiny of this paper It is enough to know

that its basic unit is the attribute-value pair,

where an attribute is a symbol (label) and a value

is a symbol or another functional description In

the sentence

(15) Ne killed her

sequence of

a value; thus the

pair is sULJshe, or

tozether with SUPJ=CAT=PRON etc Any

symbols is a path leadins to

value of SULJSNEAR is “he

The functions that access the data structure

are specifically designed to treat this type of

representation but we think that they could be

easily ceneralizcd The term “couponent” will be

used to identify the set Of naths starting fron

the sane label (radix)

The reasons of this choice are as follows:

i, addressins frapinents of information is more

easily done by followin: a path of labels than by

visiting, a unlabelied tree;

61

ii functional gremmar allows any depth of linguistic representation through the left adjunction of labels;

iii functional syntactic repesentation and lexical features (also expressed in teras of attribute= value pairs) iay be treated in a uniform way

RB The parser

used in network kept, whole set of Also the use modified to

The basic features of the ATN parser our previous experinents [1%], i-e the forwalisn and the parsing algoritina, are while the data structure and the

actions and forus have been modified

of the push-down aechanisr has been soue extent

The data structure is a list which is mainly accessed with a typical LIFO stack policy It represents a unique mewory space non splitted into registers The state savins mechanism necessary for the treatnent of non-dcterninisa is provided

by NDWLISP [2,17], the dialect of LISP in which the system is written At any point in which non-determinisu is called, the previous context,

in particular the data, is saved and only the new values are set in the current context Therefore, there is no difference between the use of the traditional resister table and this special list Since both of them are handled in the sane way This (LIFO) list contains at any point of the process the CGI, i.e the entire left context literally represented in temis of attribute~value

We pive hereafter a list in Backus notation of the Functions which access the CGH

1.Actions

as <storinzy actrions>::=

ADD pair location | ASSIGN label path

<location>::= MIL | <forma>

<label>::= any label

<pair>::= label value

<value>::= *# | <forn>

b <li§t manipulation>::=

Pusu |

Por | INSERT data iten

<data>::= any data

€item>::= <foma>

2.Forms FIID path test level dtype | FINDVAL path test level dtype | LOCATE path test level dtype

<pach>::= <labe1*+>

<test>::= T | any test

<level>::= T | CL

<dtype>::= T | ¡äD Ì L The basic storing action is

to store any inconins piece strings

ADD which is used

of structure The

(16) il cane the Jos recornized by the network

4 2

(np) Car ART + CAT N

is stored by the following actions

1 (ALD (DET *))

2, (ABD (MEAD *))

If location is NIL, the current component is

gant, otherwise the form LOCATE specifies the

path leading to the radix ta which the new pair is

to be ADNed

Relabelline of a component is done by the action

ASSICN In the sentences

(16a) 11 cane u:annria

the dos eats

(16b) Il cane e” nanriato

the doz is eaten

the NP 7il cane* will be first Labelled FOCUS or

FICSTSP Then, after havins ricocnized the verb,

the action

(ASSIGI! SUBJ (LOCATE FOCUS T CL T)) or

(ASSING OBJ (LOCATE FOCLS T CL T))

will properly classify the IP as

SUBJ or OBJ = FOCUS = DFT = IL,

ZiiEAb = 1 =2 CANE Extraction of information is done by the forms

FING, which returns a pafr, and FINDVAL, which

returns only the value of a pair LOCATE works

exactly in tie same way, but returns a pointer to

a iven radix àll tie three functions can work in

different modes They can search cither only the

current level (CL) or throush the entire list (T)

In this latter case the current level is excluded

anul, lí no further options are specified, the

lower (the nearest to the top) occurrence is

returned jnother option (<dtype) returns all the

occurrences either appenijed in a list (L} or one

by one, nonsleteruinistically (2), A third option

evaluates conditions in order to select tao

emuiponent igentified by the specifie! path

1n sentence (4) the antecedent retrievins is

perforcie! by thea forms ‘ l ta

(FINDVAL (SUBJ) (ANE (ÚQ (FIRDVAL (SUuJ NUL) TT SB)

CFINDVAL (MEA OULD)? CL T))

(EQ (EISEVAL (SULJ GEL) TT FINDVAL (HEAD GEN) TCL T))) T ND)

which searches for a subject through all the levels non detemainistically Such an NP must agree in number and wendler with the current level head, i.ee the verb(4) TẾ this expression is embedded in the function

the correct subiect(s) is (are) copied in the complement (5)

The three last actions, PUSf, POP, and INSERT, Manipulate the items in the list PUSI adds a new Ceupty) iten in front of the list The eleuents of the component being analysed (piruses or sentences) are ADDed in this top iten, which has been therefore referred to as current level, CÚP Yeaoves the curreinit topsitei and eiteds it into the new toperiten, possitly sessi min a label to the correspondin, couponent Finally LUSERT inserts an ite.:, correspondin, to «& nuv Level, somewhere back Vetveen “ites” ans the front part

of the list, and fills it with “data’

List manipulation takes place independently fron the starting or the ending of the process expressed in a subnet Thus a couponent can be POPed after the end of its reco;;nition procedure, when also its function is clarified

The are recopnizin: an ovject, for ex., can be expressed as follows

(START iiP T (COMD (FIND (SUBJ) T CL T) (POP ObJ))

(TO qi)) which aeans that if there already is a subject, the current couponent must be popped with the label OBJ

The use of the INSHRT function is primarily wotivatec by the treatment of certain relative clauses Felative pronouns are surface si rnals that trigser the enbeddin; inte a relative clause

of the currently processed cauwponent(s)

In the sentance (17)11 lisro Jells traua del đuale parbaveno The book about the plot of which we tal.sed such an @abeddiny takes vlace turcdiately «efter

*“Litro”, thas productu:

(4) An faninhoric® facility is also iaplewentec not to repeat an embedded forms with the sane arcu.ent cs the eabeadin, one

(3) Ne do not intend to suy,.est that the correct wechanisn of tracc/antecedent bindin, is the copyin, of the antecedent in the trace position A slinhtly modified version of this function visit produce tie insertion of the antecedent path, as in the orthodox functional praimar The procedure, hovever, docs not suostantially chanfe

(18) (il libro (RELCL

parlavamo))

(della trama (del quale))

The seneral rule may be formulated as follows:

“a new level labelled RELATIVECLAUSF is to be

inserted immediately after the antecedent of the

relative pronoun” Analysis of (17) will therefore

proceed as follows;

- when the relative

encountered, the foria

pronoun “quale” is

(FIND (EAD) (AND (EQ (FINDVAL (EAD GEN) TT

(FIXDVAL (DET GEN) T CL (FINDVAL (HEAD NUt) TT (FINDVAL (DET WUtl) T CL

ND) T)) ND)

T)))

(EA

T T)

returns the lower head which

gender with the determiner of “suale” (“quale” is

both masculine and feminine), i.e “libro* This

is the antecedent

- The function

agrees in nuaber and

CIUSERT RELCL (LOCATE as for FIND))

inserts a new item with label RELCL

— Ôn the sane are the Function (FOP DI-APG) enbeds

“del quale” in “della trama” and a second POP

embeds (della trama (del quale)) in the recently

inserted relative clause component

- The reco:nition of a relative

continued by a (START S ) are The control is

Finally then returned to the MP process with the

complex KP 711 libro ’as the current component

clause ts

IV ADVANTAGES

As Efficiency

The parser we have been presentin: is based on

the core alrorithn of the ATN Our modifications

affect the set of foras and actions and the data

structure The parsing alvoritha, therefore, keeps

the efficiency of traditional ATI We have already

shown that che storing of the data structure does

not present any special difference from the

traditional revisters system, even in relation to

the treatuent of non=determinisa The memory load

is, therefore, strictly a function of the Length

of the parsed seyment of the input and no overhead

fetermined by nanipulations of structures is added

as in the case of transit re;;isters

The actions and forms are equivalent to the

traditional ones, but for the Fact that iwosr of

tia cust visit the vhole left context for every

access Anyvuy tuls effect hardly balances the

setting of transit re.isters In fact, it is

worth notin, that in the iajority of coanon

sentences such accesses are very reduced, so that

(o) In this pararranh we refer to the tdeas and

saratsos [21]

63

no substantial difference exists in comparison with the traditional resister access In the discussed complex cases the access to the CCll is a known function of the lenj;th of che list, i.e of the depth of euwbedding, of the current level Within any itea search proceeds linearly as for any ordinary pattern=matching

The only substantially new fact ts the possibility of embedding the current component; this eliminates the need for backtracking, at least for soue sentences

In conclusion, it seems that if there is a difference from the traditional AT! it is in favour of the version presented here

B Generalization and modularity The set of actions and forms presented seem to provide a functional descriptioa of nany linguistic phenomena They can be reparded as liny;uistic (procedural) sencralizations, at least

on the functional sround

that linguistic phenowena can be eseribed, independently Frou the Zorialisi: that expresses then (the eYauuar), in ters of eneral operations This set of operations is open- ended and can, therefore, te increased with functions desirned for the treatment of new phenomena, as they are discussed and described Furthermore, those actions can be taken to represent nentai operations of the language user, thus providing a

This supports our claim

ia

valuable frame for psycholin;uistic experiments

It is obvious that this view strongly inclines towards the idea of parser as a collection of heuristic strategies and processes and also offers

a symmetric alternative to the HOLD hypothesis According to this hypothesis there are points in a sentence in which comprehension needs a_ heavier memory load; instead in our view an overhead of operations is sugsested Anyway the distinsuished phenomena coincide, thus keeping the inteprity of the experimental data(6)

Cc haturality Our hypothesis sees sore natural in two vays

It embeds into a non-deterministic frame sae eperations very similar to some of those designed and discussed in the deteriinistic hypotiesis (3,

4, 15, 16, 19} The result is a stron; linitation

of the effects of non-deterainisn, at Least for those cases they are designed to Creat It is interesting that startin;; fron two dpposed viewpoints couparable rasults are ovtaingd Wevertheless, as stated avove, we think tuat imposing constraints to a nonedeterainistic wouel

is sore natural than being dnposed slotal constraints by the assuaption 0Ÿ deteraiinisai In the first hypothesis, in facet, a deteriuinistic behaviour of the parser Way be uitiuately

the experiments presented by Kaplan [12, E3} and icanner &

obtained, in some points, as a result of

observation of real linguistic restrictions while

those phenomena such as ambiguity which can be

adequately treated only in a non-deterministic

frame, are not “a priori” ruled out Then, a

model such as the nonwdeterministic one, in which

there is place for the study of human heuristic

constraints, seems more attractive and natural

Qur hypothesis seems intuitively natural also

in so much as it tries to propose a “theory of

suess” During the comprehension of a sentence

guesses (CGII*s) are progressively enriched and

stored in a space of meuory During this process

errors may be done For some of them it is enough

to modify the previous guess while for others a

Teal backtracking and reanalysis is necessary

Although the distinction between the two types of

errors is unclear, it provides a valuable frane

for further research in the domain of

computational linguistics as well as

psycholinguistics In particular it seems to

đỉstinruish in the activity of sentence

couprenension a phase cf structurin;, from a nhase

of perception Errors occurrins in the former are

remedied by modifying a suess, wiile those

occurring, in the latter need backtrackins; and the

choice of another stratesy

Ve PERSPICTIVES

A more serious systematization of the proposed

functions, as well as the extension of the modei

to more and more linguistic phenomena are obvious

extensions of the present project

Another direction where investi;ation seems to

be particularly fruitfull is the relation between

syntax and semantics On one hand, the fact that

the result of the analysis is prozressively stored

in a unique space of wuenory Jo not impose special

constraints on the structure of the analyzed

Striny Ou tne other hand, uany of the presented

functions include parasueter slots for conditions

which say be Filled with any kind of test This

vodel sees, therefore, to avoid “physiological”

bounderies betveen syntax and sanantics The

stored structure can be a senantic one and = the

tests can also incorporate senantic descriptions

This sees to eventually lead to an easier

interration of the two levels ie will present

shortly [10] a first approxi.ation to a frame into

which such in interrarcion can be realized

ARNOVLERGES ENTS

1œ would lile

E{zzi, Franco

to thank Piernaolo Ueyano, Luisi Turini and Oliviero Stock who read

contriluted to it with eheizr v luaule comments

REFERENCES

(1) Aho Aw, Ullmann J.0., The theory of parsing, translation and compilins, Prentice Hall, New York, 1972

(2) Asirelli P., Lami C., MNontansero C., Pacini G., Simi !⁄., Turini Fe, ivecuaAeLISP Seference lianual, NT C75-13, IEL C’m, Pisa, 1975

(3) Berwick B.C., Constraints nad Proceedings of the Cambridge MA,15-17

Weinbers, AeSe, Syntactic efficient Parsability, in 2lst Annual weetiny of ACL, jJune 1935, pp.Llv=l22 (4) serwick 0.2, Á deterwinistic parser with broad coverace, in Proceédinys of the Kkighth International Joint Conference on artificial Intellipzence, Karlsruhe @-12 August 1983, pp 710-712

(5) Cappelli A., Ferrari G., :oretti Le, Prodanof I., Stock O., Parsing an Italian Text with an ATN parser, NT ILC Cin, Pisa, 1976

(6) Cappelli A., Ferrari G., tboretti L., Prodanof I., Stock O., Automatic Analysis of Italian, in Proceedings of the AISB-=l Conference on Artificial Intellirence, Amsterdam, 1-4 July,

1980

(7) Canpelli aA., Perrari G., toretti L., Prodanof I., Stock G., Il trattamento di alcuni fenomeni anaforici wediante un aATN, in Atti del seninario “Sul11“anaforn“;, Firenze 14-16/12/1931, pp‹275-249

(5) Cappelli A., Ferrari G., ::oretti La, Prodanof I., Stock O., Costruzione, sperinentazione ed estensione di un ATi cowe wodello di analisi del linsuasyio naturale, in Picerche di Psicolo;ia, VII, 25, 1963, jp 159-104

(9) Cappelli A., Ferrari 6G, :oretti L.; Prodanof I., Towards an intesrated wuodel of sentence comprehension, in A.Cappelli (ed.) Kesearch in Natural Languase Processin;, in Italy (1951), Giardini, Pisa, 1943, pp 45-56

(10) Cappelli A., Ferrari C., ‘oretti L., Prodanof

te, A framework for interratin syntax and semantics, in Lara & & Guida 6G (cds.) Natural Lancuage Processinn; North Holland, (to appear)

(11) Ferrari C., Stock G., Stratesie selection for

an ATi! Syntactic Parser, in Proceecinys of the Sth Annual weetins: of ACL, Philadelphia, June L&=22 1930, pp 113-115

(12) Kaplan &., Augmented Transition tetworks as

psychological Models for Sentence

Comprehension, Artificial Intellisence 3, 1972,

ĐPs 77-100

(13) Kaplan &., On process !iodels for sentence

Analysis, in Norman & Rumelharr 2b (eds.)

Explorations in Cognition,

1975, pp- 117-135

Freeman, S.Francisco

(14) Xay !1., Funetional Grammar, in Proceedings of

the 5th !leeting of the Lerkeley Linguistic

Society, ferkeley 1973, pp 142~158

(15) Marcus H., A Theory of Syntactic Recognition

for Natural Lansjuacze, NIT Press, Canbridsre iA,

1950

(16) !iarcus !⁄2P., Uindle be, Fleck l2, D^Theory:

Talkiur about Tail:ins; about Trees, in

Proceedines of the 21st Annual iieeting of ACL,

Canbridge NA, 15-17 June 1903, pp 129-138

(17) ;ontanrero C., Pacini G., Turini ¥., WD-LISP

Reference sianual, ST C76<3, IfI CHP, Pisa,

1975

(18) Prodanof I., Ferrari G., ktevisinr and ATI

parser, in COLI!Gÿ2, Proceedinss of the 9th

65

Conference on Prague, 5-10 July

International Linguistics, 101-165

Coaputational

1532, pn

(19) Shipman W.P., liarcus i , Towards mininal data structure for deterministic parsing, in Proceedings of the th International Joint Conference on Artificial Intellizence, Tokyo, August 20~23 1979, pp.81l5—-s17

(20) Stock O., ATUSYS: un sistema per l*analisi gramuaticale autowatica delle lLin:ue naturall,

NI B76=29, IEI CHR, Pisa, 1976

(21) Wanner £., iiaratsos tie, An ATS comprehension, in i alle, J.uresnan, G siiller (eds) Linguistic Theory and Psycholo,ical reality, “IT Fress, Caubridsue iia,

1975, py.11l9-161

approach to

(22) Vinoprad f., Laneuace as a Cognitive Process Syntax, Addisonslesley, seading id, 1903 (23) Woods W., Transition iletwork Cramears for Natural Lan uae Analysis, im CACi: 13-10, 197¢,