1. Trang chủ
  2. » Luận Văn - Báo Cáo

Báo cáo khoa học: "A Knowledge-free Method for Capitalized Word Disambiguation" doc

8 336 0
Tài liệu đã được kiểm tra trùng lặp

Đang tải... (xem toàn văn)

THÔNG TIN TÀI LIỆU

Thông tin cơ bản

Định dạng
Số trang 8
Dung lượng 808,21 KB

Các công cụ chuyển đổi và chỉnh sửa cho tài liệu này

Nội dung

Such words can act as proper names or can be j u s t capitalized vari- ants of common words.. Such ambiguous positions include the first word in a sentence, words in all-capitalized titl

Trang 1

A Knowledge-free M e t h o d for Capitalized Word Disambiguation

A n d r e i M i k h e e v *

H a r l e q u i n L t d , L i s m o r e H o u s e , 127 G e o r g e S t r e e t , E d i n b u r g h E H 7 2 4 J N , U K

mikheev@harlequin, c o uk

A b s t r a c t

In this paper we present an a p p r o a c h to the dis-

ambiguation of capitalized words when t h e y are

used in the positions where capitalization is ex-

pected, such as the first word in a sentence or

after a period, quotes, etc Such words can act

as proper names or can be j u s t capitalized vari-

ants of common words T h e main feature of

our approach is that it uses a m i n i m u m of pre-

built resources and tries to dynamically infer

the disambiguation clues from the entire docu-

ment The approach was t h o r o u g h l y tested and

achieved a b o u t 98.5% accuracy on unseen texts

from The New York Times 1996 corpus

1 I n t r o d u c t i o n

Disambiguation of capitalized words in mixed-

case texts has hardly received much attention

in the natural language processing and infor-

mation retrieval communities, b u t in fact it

plays an important role in m a n y tasks Cap-

italized words usually denote p r o p e r names -

names of organizations, locations, people, arti-

facts, etc - b u t there are also other positions in

the text where capitalization is expected Such

ambiguous positions include the first word in

a sentence, words in all-capitalized titles or ta-

ble entries, a capitalized word after a colon or

open quote, the first capitalized word in a list-

entry, etc Capitalized words in these and some

other positions present a case of a m b i g u i t y -

they can stand for proper names as in "White

later said .", or they can be j u s t capitalized

common words as in "White elephants are ."

T h u s the disambiguation of capitalized words in

the ambiguous positions leads to the identifica-

tion of proper names I and in this p a p e r we will

* Also at HCRC, U n i v e r s i t y of E d i n b u r g h

1This is n o t e n t i r e l y t r u e - a d j e c t i v e s d e r i v e d f r o m lo-

c a t i o n s s u c h as A m e r i c a n , F r e n c h , etc., are always w r i t -

use these two terms interchangeably Note t h a t this task, does not involve the classification of proper names into semantic categories (person, organization, location, etc.) which is the objec- tive of the N a m e d E n t i t y Recognition task Many researchers observed t h a t c o m m o n l y used u p p e r / l o w e r case normalization does not necessarily help d o c u m e n t retrieval Church in (Church, 1995) among other simple text nor- malization techniques studied the effect of case normalization for different words and showed that " sometimes case variants refer to the same thing (hurricane and Hurricane), some- times t h e y refer to different things (continental

and Continental) and sometimes they d o n ' t re- fer to much of anything (e.g anytime and Any- time)." Obviously these differences are due to the fact that some capitalized words s t a n d for proper names (such as Continental- the n a m e

of an airline) and some don't

P r o p e r names are the main concern of the

N a m e d E n t i t y Recognition s u b t a s k (Chinchor, 1998) of Information Extraction There the dis-

a m b i g u a t i o n of the first word of a sentence (and

in other ambiguous positions) is one of the cen- tral problems For instance, the word "Black"

in the sentence-initial position can s t a n d for

a person's s u r n a m e b u t can also refer to the colour Even in multi-word capitalized phrases the first word can belong to the rest of the phrase or can be j u s t an external modifier In the sentence "Daily, Mason and P a r t n e r s lost their court case" it is clear t h a t "Daily, Mason and Partners" is the name of a company In the sentence "Unfortunately, Mason and P a r t n e r s lost their court case" the name of the c o m p a n y does not involve the word "unfortunately", b u t

t e n c a p i t a l i z e d b u t in fact c a n s t a n d for a n a d j e c t i v e

(American president) as well as a p r o p e r n o u n (he was

an American)

Trang 2

the word "Daily" is just as c o m m o n a word as

"unfortunately"

Identification of proper names is also impor-

tant in Machine Translation because normally

proper names should be transliterated (i.e pho-

netically translated) rather t h a n properly (se-

mantically) translated In confidential texts,

such as medical records, proper names must be

identified and removed before making such texts

available to u n a u t h o r i z e d people And in gen-

eral, most of t h e tasks which involve different

kinds of text analysis will benefit from the ro-

bust disambiguation of capitalized words into

proper names and capitalized c o m m o n words

Despite the obvious importance of this prob-

lem, it was always considered p a r t of larger

tasks and, t o the authors' knowledge, was not

studied closely with full attention In the part-

of-speech tagging field, the disambiguation of

capitalized words is t r e a t e d similarly to the

disambiguation of c o m m o n words However,

as Church (1988) rightly pointed out " P r o p e r

nouns and capitalized words are particularly

problematic: some capitalized words are p r o p e r

nouns and some are not Estimates from t h e

Brown Corpus can be misleading For exam-

ple, the capitalized word "Acts" is found twice

in Brown Corpus, b o t h times as a proper n o u n

(in a title) It would be misleading to infer

from this evidence t h a t the word "Acts" is al-

ways a proper noun." Church t h e n proposed to

include only high frequency capitalized words

in the lexicon and also label words as proper

nouns if t h e y are "adjacent to" other capital-

ized words For the rest of capitalized c o m m o n

words he suggested t h a t a small probability of

proper noun interpretation should be assumed

and t h e n one should hope t h a t the surrounding

context will help to make the right assignment

This approach is successful for some cases but,

as we pointed out above, a sentence-initial cap-

italized word which is adjacent to other capital-

ized words is not necessarily a part of a proper

name, and also m a n y c o m m o n nouns and plural

nouns can be used as proper names (e.g Rid-

ers) and their contextual expectations are not

too different from their usual parts of speech

In the Information Extraction field the dis-

ambiguation of capitalized words in the am-

biguous positions was always tightly linked to

the classification of the proper names into se-

mantic classes such as person name, location,

c o m p a n y name, etc and to the resolution of coreference between the identified and classi- fied proper names This gave rise to the meth- ods which aim at these tasks simultaneously (Mani&MacMillan, 1995) describe a m e t h o d

of using contextual clues such as appositives

physician") and felicity conditions for identify- ing names T h e contextual clues themselves are

t h e n t a p p e d for d a t a concerning the referents

of the names T h e advantage of this approach

is t h a t these contextual clues not only indicate

w h e t h e r a capitalized word is a proper name,

b u t t h e y also determine its semantic class T h e disadvantage of this m e t h o d is in the cost and difficulty of building a wide-coverage set of con-

t e x t u a l clues and the d e p e n d e n c e of these con-

t e x t u a l clues on the d o m a i n a n d text genre Contextual clues are very sensitive to the spe- cific lexical and syntactic constructions and the clues developed for the news-wire texts are not useful for legal or medical texts

In this paper we present a novel approach to

t h e problem of capitalized word disambiguation

T h e m a i n feature of our approach is t h a t it uses

a m i n i m u m of pre-built resources and tries to

d y n a m i c a l l y infer the disambiguation clues from the entire d o c u m e n t u n d e r processing This makes our approach d o m a i n and genre inde-

p e n d e n t and thus inexpensive to apply w h e n dealing with unrestricted texts This approach was used in a n a m e d entity recognition s y s t e m (Mikheev et al., 1998) where it proved to be one of the key factors in the system achieving a nearly h u m a n performance in the 7th Message

U n d e r s t a n d i n g Conference (MUC'7) evaluation (Chinchor, 1998)

2 B o t t o m - L i n e P e r f o r m a n c e

In general, the disambiguation of capitalized words in the mixed case texts d o e s n ' t seem to

be too difficult: if a word is capitalized in an un- ambiguous position, e.g., not after a period or

o t h e r p u n c t u a t i o n which might require the fol- lowing word to be capitalized (such as quotes or brackets), it is a proper n a m e or p a r t of a multi- word proper name However, when a capitalized word is used in a position where it is expected

to be capitalized, for instance, after a period or

in a title, our task is to decide w h e t h e r it acts

Trang 3

Total Words

P r o p e r Names

C o m m o n Words

All Words tokens types

2,677 665

826 339 1,851 326

Known Words tokens types

2,012 384

1,841 316

Unknown Words tokens types

Table 1: Distribution of capitalized w o r d - t o k e n s / w o r d - t y p e s in the ambiguous positions

as a proper name or as the expected capitalized

c o m m o n word

T h e first obvious strategy for deciding

w h e t h e r a capitalized word in an ambiguous po-

sition is a proper name or not is to apply lexi-

con lookup (possibly enhanced with a morpho-

logical word guesser, e.g., (Mikheev, 1997)) and

mark as proper names the words which are not

listed in the lexicon of c o m m o n words Let us

investigate this strategy in more detail: In our

experiments we used a corpus of 100 d o c u m e n t s

(64,337 words) from The New York Times 1996

This corpus was balanced to represent different

domains and was used for the formal test run

of the 7th Message U n d e r s t a n d i n g Conference

( M U C ' 7 ) (Chinchor, 1998) in the N a m e d En-

tity Recognition task

First we ran a simple zoner which identi-

fied ambiguous positions for capitalized words -

capitalized words after a period, quotes, colon,

semicolon, in all-capital sentences and titles

and in the beginnings of itemized list entries

T h e 64,337-word corpus contained 2,677 cap-

italized words in ambiguous positions, out of

which 2,012 were listed in the lexicon of En-

glish c o m m o n words Ten c o m m o n words were

not listed in the lexicon and not guessed by our

morphological guesser: "Forecasters", "Bench-

mark", " E e v e r y b o d y " , "Liftoff", "Download-

ing", "Pretax", "Hailing", "Birdbrain", "Opt-

ing" and "Standalone" In all our experiments

we did not try to disambiguate b e t w e e n singu-

• lar and plural proper names and we also did

not count as an error the adjectival reading of

words which are always w r i t t e n capitalized (e.g

American, Russian, Okinawian, etc.) T h e dis-

t r i b u t i o n of proper names among the ambiguous

capitalized words is shown in Table 1

Table 1 allows one to e s t i m a t e the perfor-

m a n c e of the lexicon lookup s t r a t e g y which we

take as the bottom-line First, using this strat-

egy we would wrongly assign the ten c o m m o n

words which were not listed in the lexicon More damaging is the biind assignment of the com- mon word category to the words listed in the lexicon: out of 2,012 known word-tokens 171 actually were used as proper names This in to- tal would give us 181 errors o u t of 2,677 tries

- a b o u t a 6.76% misclassification error on capi- talized word-tokens in the ambiguous positions The lexicon lookup strategy can be enhanced

by accounting for the i m m e d i a t e context of the capitalized words in question However, cap- italized words in the ambiguous positions are not easily d i s a m b i g u a t e d by their surrounding part-of-speech context as a t t e m p t e d by part-of- speech taggers For instance, many surnames are at the same time nouns or plural nouns in English and thus in b o t h variants can be fol- lowed by a p a s t tense verb Capitalized words

in the phrases Sails rose or Feeling him-

s e l l , can easily be interpreted either way and only knowledge of semantics disallows the plural noun interpretation of Stars can read

Another challenge is to decide whether the first capitalized word belongs to the group of the following p r o p e r nouns or is an external modifier and therefore not a proper noun For instance,

All American B a n k is a single phrase b u t in All State Police the word "All" is an external mod- ifier and can be safely decapitalized One might argue that a part-of-speech tagger can c a p t u r e that in the first case the word "All" modified a singular p r o p e r noun ( " B a n k " ) and hence is not grammatical as an external modifier and in the second case it is a grammatical external modi- fier since it modifies a plural proper noun ("Po- lice") b u t a simple counter-example - All Amer- ican Games - defeats this line of reasoning The third challenge is of a more local nature

- it reflects a capitalization convention a d o p t e d

by the author For instance, words which re- flect the o c c u p a t i o n of a person can b e used in

an honorific m o d e e.g "Chairman Mao" vs

Trang 4

"ATT chairman Smith" or "Astronaut Mario

Runko" vs "astronaut Mario Runko" W h e n

such a phrase opens a sentence, looking at the

sentence only, even a h u m a n classifier has trou-

bles in making a decision

To evaluate the performance of part-of-speech

taggers on the p r o p e r - n o u n identification task

we ran an H M M trigram tagger (Mikheev, 1997)

and the Brill tagger (Brill,.1995) on our cor-

pus B o t h taggers used the P e n n Treebank tag-

set and were trained on the Wall Street Jour-

nal corpus (Marcus et al., 1993) Since for our

task the mismatch between plural p r o p e r noun

(NNPS) and singular proper n o u n (NNP) was not

i m p o r t a n t we did not count this as an error De-

pending on the smoothing technique, the H M M

tagger performed in the range of 5.3%-4.5% of

the misclassification error on capitalized com-

mon words in the ambiguous positions, and the

Brill tagger showed a similar p a t t e r n when we

varied the lexicon acquisition heuristics

T h e taggers handled the cases when a poten-

tial adjective was followed by a verb or adverb

( "Golden added ") well b u t t h e y got confused

with a potential noun followed by a verb or

adverb ( "Butler was " vs "Safety was "),

p r o b a b l y because the taggers could not distin-

guish between concrete and mass nouns Not

surprisingly the taggers did not do well on po-

tential plural nouns and gerunds - none of t h e m

were assigned as a proper noun T h e taggers

also could not handle the case when a poten-

tial noun or adjective was followed by another

capitalized word ("General Accounting Office")

well In general, when the taggers did not have

strong lexical preferences, apart from several

obvious cases they tended to assign a c o m m o n

word category to known capitalized words in the

ambiguous positions and the p e r f o r m a n c e of the

part-of-speech tagging approach was only a b o u t

2% superior to the simple b o t t o m - l i n e strategy

3 O u r K n o w l e d g e - F r e e M e t h o d

As we discussed above, the b a d news (well, not

really news) is that virtually any c o m m o n word

can potentially act as a proper name or part of

a multi-word proper name Fortunately, there

is good news too: ambiguous things are usu-

ally u n a m b i g u o u s l y introduced at least once in

the text unless they are part of c o m m o n knowl-

edge p r e s u p p o s e d to be known by the readers

This is an observation which can be applied to

a broader class of tasks For example, people are often referred to by their surnames (e.g

"Black") b u t usually i n t r o d u c e d at least once

in the text either with their first name ( " J o h n Black") or with their title/profession affiliation ("Mr Black", "President B u s h " ) and it is only when their names are c o m m o n knowledge t h a t they don't need an i n t r o d u c t i o n ( e.g "Castro",

"Gorbachev")

In the case of proper n a m e identification we are not concerned with the semantic class of a name (e.g w h e t h e r it is a person name or loca- tion) b u t we simply want to distinguish w h e t h e r this word in this particular occurrence acts as

a proper name or p a r t of a multi-word p r o p e r name If we restrict our scope only to a single sentence, we might find t h a t there is j u s t not enough information to make a confident deci- sion For instance, Riders in the sentence "Rid- ers said later " is equally likely to be a p r o p e r

noun, a plural p r o p e r n o u n or a plural com- mon noun b u t if in the same text we find " J o h n Riders" this sharply increases the proper n o u n interpretation and conversely if we find "many riders" this suggests the plural noun interpre- tation Thus our suggestion is to look at the unambiguous usage of the words in question in the entire document

3.1 T h e S e q u e n c e S t r a t e g y Our first s t r a t e g y for the disambiguation of cap- italized words in a m b i g u o u s positions is to ex- plore sequences of p r o p e r nouns in unambigu- ous positions We call it the Sequence Strategy

The rationale b e h i n d this is t h a t if we detect a phrase of two or more capitalized words and this phrase starts from an u n a m b i g u o u s position we can be reasonably confident t h a t even when the same phrase starts from an unreliable position all its words still have to b e g r o u p e d together and hence are p r o p e r nouns Moreover, this ap- plies not just to the exact replication o f such a phrase b u t to any partial ordering of its words of size two or more preserving their sequence For instance, if we detect a phrase Rocket Systems Development Co in the middle of a sentence,

we can mark words in the sub-phrases Rocket Systems, Rocket Systems Co., Rocket Co., Sys- terns Development, etc as proper nouns even if

they occur at the beginning of a sentence or in other ambiguous positions A span of capital-

Trang 5

Proper Names Common Words Total

All Ambiguous

Disambiguated +

Sequence Strategy +

Single Word +

Assignment

Stop-List -t-

Assignment

All Words tokens types

826 339

795

1

62

0

Known Words All Words tokens types tokens types

171 68 1,851 326

54 1,568 218

148

3

Known Words All Words tokens types tokens types

1,841 316 2,677 665

1,563 213

510

1

0

0

70

0

1,265 143

316 140

192 108

43 1,270

298 70

0

2,363 534

1,780 340

298 70

0

Table 2: Disambiguated capitalized w o r d - t o k e n s / t y p e s in the ambiguous positions

ized words can also include lower-cased words of

length three or shorter This allows us to cap-

ture phrases like A ~ M, T h e P h a n t o m of the

Opera., etc We generate p a r t i a l orders from

such phrases in a similar way b u t insist t h a t ev-

ery generated sub-phrase should s t a r t and end

with a capitalized word

To make the Sequence S t r a t e g y robust to po-

tential capitalization errors in the d o c u m e n t we

also use a set of negative evidence This set is

essentially a set of all lower-cased words of the

document with their following words (bigrams)

We don't a t t e m p t here to build longer sequences

and their partial orders because we cannot in

general restrict the scope of dependencies in

such sequences The negative evidence is t h e n

used together with the positive evidence of the

Sequence Strategy and block the proper name

assignment when controversy is found For in-

stance, if in a document the s y s t e m detects a

capitalized phrase "The P r e s i d e n t " in an un-

ambiguous position, t h e n it will be assigned as

a proper name even if found in ambiguous po-

sitions in the same document To be more pre-

cise the m e t h o d will assign the word "The" as a

proper noun since it should be grouped together

with the word "President" into a single proper

name However, if in the same d o c u m e n t the

system detects an alternative evidence e.g "the

P r e s i d e n t " or "the p r e s i d e n t " - it t h e n blocks

such assignment as unsafe

T h e Sequence Strategy strategy is e x t r e m e l y useful when dealing with names of organizations since m a n y of t h e m are multi-word phrases com- posed from c o m m o n words And indeed, as is shown in Table 2, the precision of this strat- egy was 100% a n d t h e recall a b o u t 7.5%: out

of 826 proper names in ambiguous positions, 62 were marked and all of t h e m were marked cor- rectly If we concentrate only on difficult cases when proper names are at the same time com-

m o n words of English, the recall of the Sequence Strategy rises to 18.7%: out of 171 c o m m o n words which acted as proper names 32 were cor- rectly marked Among such words were "News" from "News Corp.", "Rocket" from "Rocket Systems Co.", "Coast" from "Coast G u a r d " a n d

"To" from "To B Super"

T h e Sequence Strategy is accurate, b u t it cov- ers only a p a r t of potential proper names in ambiguous positions and at the same time it does not cover cases when capitalized words do not act as proper names For this purpose we developed a n o t h e r strategy which also uses in-

f o r m a t i o n from the entire document We call

this s t r a t e g y Single W o r d A s s i g n m e n t , a n d it

can be s u m m a r i z e d as follows: if we detect a word which in the current d o c u m e n t is seen capitalized in an unambiguous position a n d at the same time it is not used lower-cased, this word in this particular d o c u m e n t , even when

Trang 6

used capitalized in ambiguous positions, is very

likely to s t a n d for a proper name as well And

conversely, if we detect a word which in the

current d o c u m e n t is used only lower-cased in

u n a m b i g u o u s positions, it is extremely unlikely

t h a t this word will act as a proper n a m e in an

ambiguous position and thus, such a word can

be marked as a common word T h e only consid-

eration here should be made for high frequency

sentence-initial words which do not n o r m a l l y

act as proper names: even if such a word is

observed in a d o c u m e n t only as a proper n a m e

(usually as p a r t of a multi-word proper name),

it is still not safe to mark it as a proper n a m e in

ambiguous positions Note, however, t h a t these

words can be still marked as proper names (or

r a t h e r as parts of proper multi-word names) by

the Sequence Strategy To build such list of

stop-words we ran the Sequence S t r a t e g y a n d

Single Word Assignment on the Brown Corpus

(Francis&Kucera, 1982), and reliably collected

100 most frequent sentence-initial words

Table 2 shows the success of the Single Word

Assignment strategy: it marked 511 proper

names from which 510 were marked correctly,

a n d it marked 1,273 c o m m o n words from which

1,270 were marked correctly T h e only word

which was incorrectly marked as a proper n a m e

was the word "Insurance" in "Insurance com-

p a n y ." because in the same d o c u m e n t there

was a proper phrase "China-Pacific Insurance

Co." a n d no lower-cased occurrences of the

word "insurance" were found T h e three words

incorrectly marked as c o m m o n words were:

"Defence" in "Defence officials ", "Trade" in

"Trade Representation office " a n d "Satellite"

in "Satellite Business News" Five out of ten

words which were not listed in the lexicon ( "Pre-

t a x " , " B e n c h m a r k " , " L i f t o f f ' , "Downloading"

a n d " S t a n d a l o n e " ) were correctly m a r k e d as

c o m m o n words because t h e y were found to ex-

ist lower-cased in the text In general the error

rate of the assignment by this m e t h o d was 4 out

of 1,784 which is less t h a n 0.02% It is interest-

ing to m e n t i o n t h a t when we ran Single Word

Assignment w i t h o u t the stop-list, it incorrectly

marked as proper names only three e x t r a com-

m o n words ("For", "People" and " M O R E " )

3.3 T a k i n g C a r e o f t h e R e s t

After Single Word Assignment we applied a sim-

ple s t r a t e g y of marking as c o m m o n words all

unassigned words which were found in the stop- list of the most frequent sentence-initial words This gave us no errors a n d covered extra 298

c o m m o n words In fact, we could use this strat- egy before Single Word Assignment, since the words from the stop-list are not marked at t h a t point anyway Note, however, t h a t the Sequence Strategy still has to be applied prior to the stop- list assignment A m o n g the words which failed

to be assigned by either of our strategies were

243 proper names, b u t only 30 of t h e m were

in fact ambiguous, since t h e y were listed in the lexicon of c o m m o n words So at this point we marked as proper names all unassigned words which were not listed in the lexicon of c o m m o n words This gave us 223 correct assignments

a n d 5 incorrect ones - the r e m a i n i n g five out of these ten c o m m o n words which were not listed

in the lexicon So, in total, by the combination

of the described m e t h o d s we achieved a

precision of c o r r e c t l y - a s s i g n e d 2 3 6 3 - - 99.62%

a l l _ a s s i g n e d - - 2 3 6 3 + 9 -

t o t a l _ a m b i g u o u s - - 2 6 7 7 - -

Now we have to decide w h a t to do with the re- maining 305 words which failed to be assigned

A m o n g such words there are 275 c o m m o n words

a n d 30 proper names, so if we simply m a r k all these words as c o m m o n words we will increase our recall to 100% w i t h some decrease in pre- cision - from 99.62% down to 98.54% A m o n g the unclassified proper names there were a few which could be dealt by a part-of-speech tag- get: "Gray, chief ", "Gray said ", "Bill Lat- tanzi ", "Bill Wade ", "Bill Gates ", "Burns , an " a n d " Golden a d d e d " A n o t h e r four un- classified proper names were capitalized words which followed the "U.S." abbreviation e.g

"U.S Supreme C o u r t " This is a difficult case even for sentence b o u n d a r y d i s a m b i g u a t i o n sys- terns ((Mikheev, 1998), (Palmer&Hearst, 1997)

a n d ( R e y n a r & R a t n a p a r k h i , 1997)) which are built for exactly t h a t purpose, i.e., to decide whether a capitalized word which follows an ab- breviation is a t t a c h e d to it or w h e t h e r there is a sentence b o u n d a r y between them T h e "U.S." abbreviation is one of t h e most difficult ones because it can be as often seen at the end of

a sentence as in the beginning of multi-word proper names A n o t h e r nine unclassified proper names were stable phrases like "Foreign Min- ister", "Prime Minister", "Congressional Re- publicans", "Holy Grail", etc m e n t i o n e d j u s t

Trang 7

once in a document And, finally, a b o u t seven

or eight unclassified proper names were diffi-

cult to account for at all e.g "Sate-owned"

or "Freeman Zhang" Some of the above men-

tioned proper names could b e resolved if we ac-

cumulate multi-word proper names across sev-

eral documents, i.e., we can use information

from one d o c u m e n t when we deal with another

This can be seen as an extension to our Se-

quence Strategy with the only difference that

the proper noun sequences have to b e taken not

only from the current d o c u m e n t b u t from the

cache m e m o r y and all multi-word proper names

identified in a d o c u m e n t are to be a p p e n d e d

to that cache W h e n we tried this strategy on

our test corpus we were able to correctly assign

14 out of 30 remaining p r o p e r names which in-

creased the s y s t e m ' s precision on the corpus to

99.13% with 100% recall

4 D i s c u s s i o n

In this paper we presented an approach to the

disambiguation of capitalized c o m m o n words

when they are used in positions where capi-

talization is expected Such words can act as

proper names or can be j u s t capitalized variants

of common words T h e main feature of our ap-

proach is that it uses a m i n i m u m of pre-built

resources - we use only a list of common words

of English and a list of the most frequent words

which a p p e a r in the sentence-stating positions

Both of these lists were acquired without any

human intervention To c o m p e n s a t e for the lack

of pre-acquired knowledge, the system tries to

infer disambiguation clues from the entire doc-

ument itself This makes our approach domain

independent and closely targeted to each docu-

ment Initially our m e t h o d was developed using

the training d a t a of the MUC-7 evaluation and

tested on the withheld test-set as described in

this paper We then applied it to the Brown

Corpus and achieved similar results with degra-

dation of only 0.7% in precision, mostly due to

the text zoning errors and unknown words We

deliberately s h a p e d our approach so it does not

rely on pre-compiled statistics but rather acts

by analogy This is because the most interest-

ing events are inherently infrequent and, hence,

are difficult to collect reliable statistics for, and

at the same time pre-compiled statistics would

be s m o o t h e d across multiple d o c u m e n t s rather

t h a n targeted to a specific d o c u m e n t The main strategy of our a p p r o a c h is to scan the entire d o c u m e n t for u n a m b i g u o u s usages of words which have to be disambiguated T h e fact that the pre-built resources are used only

at the latest stages of processing (Stop-List Assignment and Lexicon L o o k u p Assignment) ensures that the system can handle u n k n o w n words and disambiguate even very implausible proper names For instance, it correctly as- signed five out of ten u n k n o w n c o m m o n words

A m o n g the difficult cases resolved by the sys- tem were a multi-word proper n a m e "To B Su- per" where b o t h "To" and "Super" were cor- rectly identified as proper nouns and a multi- word p r o p e r name "The U p d a t e " where "The" was correctly identified as p a r t of the maga- zine name B o t h "To" and "The" were listed

in the stop-list and therefore were very implau- sible to classify as proper nouns b u t neverthe- less the s y s t e m handled t h e m correctly In its generic configuration the s y s t e m achieved pre- cision of 99.62% with recall of 88.7% and preci- sion 98.54% with 100% recall W h e n we en- hanced the system with a multi-word p r o p e r name cache m e m o r y the p e r f o r m a n c e improved

to 99.13% precision with 100% recall This is

a statistically significant improvement against the b o t t o m - l i n e performance which fared a b o u t 94% precision with 100% recall

One of the key factors to the success in the

p r o p o s e d m e t h o d is an accurate zoning of the documents Since our m e t h o d relies on the cap- italization in u n a m b i g u o u s positions - such po- sitions should be r o b u s t l y identified In the general case this is not too difficult b u t one should take care of titles, q u o t e d speech and list entries - otherwise if t r e a t e d as ordinary text they can provide false candidates for cap- italization Our m e t h o d in general is not too sensitive to the capitalization errors: the Se- quence S t r a t e g y is complimented with the neg- ative evidence This together with the fact that

it is rare when several words a p p e a r by mistake more than once makes this s t r a t e g y robust T h e Single Word Assignment s t r a t e g y uses the stop list which includes the most frequent c o m m o n words This screens out m a n y potential errors One notable difficulty for the Single Word As- signment represent words which denote profes- sion/title affiliations These words modifying

Trang 8

a person n a m e might require capitalization -

"Sheriff John Smith", but in the same docu-

ment they can appear lower-cased - "the sher-

iff" W h e n the capitalized variant occurs only

as sentence initial our m e t h o d predicts t h a t it

should be decapitalized This, however, is an

extremely difficult case even for h u m a n index-

ers - some writers t e n d to use certain profes-

sions such as Sheriff, Governor, A s t r o n a u t , etc.,

as honorific affiliations and others t e n d to do

otherwise This is a generally difficult case for

Single Word Assignment - when a word is used

as a proper n a m e a n d as a c o m m o n word in

the same d o c u m e n t , a n d especially when one of

these usages occurs only in an ambiguous posi-

tion For instance, in a d o c u m e n t a b o u t steel

the only occurrence of "Steel C o m p a n y " hap-

pened to s t a r t a sentence This lead to an er-

roneous assignment of the word "Steel" as com-

mon noun A n o t h e r example: in a d o c u m e n t

a b o u t "the Acting J u d g e " , the word "acting"

in a sentence "Acting on behalf " was wrongly

classified as a proper name

The described approach is very easy to imple-

ment and it does not require training or installa-

tion of other software T h e system can be used

as it is and, by implementing the cache mem-

ory of multi-word proper names, it can be tar-

geted to a specific domain T h e system can also

be used as a pre-processor to a part-of-speech

tagger or a sentence b o u n d a r y disambiguation

program which can t r y to apply more sophisti-

cated m e t h o d s to unresolved capitalized words

In fact, as a by-product of its performance,

our system d i s a m b i g u a t e d a b o u t 17% (9 out of

60) of ambiguous sentence boundaries when an

abbreviation was followed by a capitalized word

Apart from collecting an extensive cache of

multi-word proper names, another useful strat-

egy which we are going to test in the f u t u r e is

to collect a list of c o m m o n words which, at the

beginning of a sentence, act most frequently as

proper names a n d to use such a list in a simi-

lar fashion to the list of stop-words Such a list

can be collected completely a u t o m a t i c a l l y b u t

this requires a corpus or corpora much larger

t h a n the Brown Corpus because the relevant

sentences are r a t h e r infrequent We are also

planning to investigate the sensitivity of our

m e t h o d to the d o c u m e n t size in more detail

R e f e r e n c e s Brill E 1995 "Transformation-based error-driven learning and natural language parsing: a case study in part-of-speech tagging" In Computa- tional Linguistics 21 (4), pp 543-565

N Chinchor 1 9 9 8 Overview of MUC-7 In

Seventh Message Understanding Conference (MUC- 7) : Proceedings of a Conference held in Fairfax, VA, April 29-May 1, 1998

www muc s a i c com/muc_7_proceedings/overwiew, html

K Church 1995 "One Term Or Two?" In Pro- ceedings of the 18th Annual Internationals ACM SIGIR Conference on Research and Development

in Information Retrieval (SIGIR'95), Seattle

K Church 1988 A Stochastic parts program and noun-phrase parser for unrestricted text In Pro- ceedings of the Second A CL Conference on Ap- plied Natural Language Processing (ANLP'88),

Austin, Texas

W Francis and H Kucera 1982 Frequency Analysis

of English Usage Boston MA: Houghton Mifflin

D D Palmer and M A Hearst 1997 Adaptive Mul- tilingual Sentence Boundary Disambiguation In

Computational Linguistics, 23 (2), pp 241-269

I Mani and T.R MacMillan 1 9 9 5 Identifying Unknown Proper Names in Newswire Text In

B Boguraev and J Pustejovsky, eds., Corpus Processing for Lexical Acquisition, MIT Press

M Marcus, M.A Marcinkiewicz, and B Santorini

1993 Building a Large Annotated Corpus of En- glish: The Penn Treebank In Computational Lin- guistics, vol 19(2), ACL

A Mikheev 1 9 9 8 "Feature Lattices for Maxi- mum Entropy Modelling" In Proceedings of the 36th Conference of the Association for Compu- tational Linguistics (A CL/COLING'98), pp 848-

854 Montreal, Quebec

A Mikheev 1997 "Automatic Rule Induction for Unknown Word Guessing." In Computational Linguistics 23 (3), pp 405-423

A Mikheev 1997 "LT POS - the LTG part of speech tagger." Language Tech- nology Group, University of Edinburgh

www Itg ed ac uk/software/pos

A Mikheev, C Grover and M Moens 1998 De- scription of the L T G system used for M U C - 7

In Seventh Message Understanding Confer- ence (MUC-7): Proceedings of a Conference held in Fairfax, VA, April 29-May I, 1998

www.muc, s a i c com/muc_7_proceedings/ltg- muc7 ps

J C Reynar and A Ratnaparkhi 1997 A Max-

i m u m Entropy Approach to Identifying Sentence Boundaries In Proceedings of the Fifth A CL Con- ference on Applied Natural Language Processing (ANLP'97), Washington D.C., ACL

Ngày đăng: 17/03/2014, 07:20

TÀI LIỆU CÙNG NGƯỜI DÙNG

TÀI LIỆU LIÊN QUAN