Practical Partial Parser for Biomedical

Literature Summarization

Yasunori Yamamoto

and Toshihisa Takagi

Department of Computer Science, University of Tokyo, Tokyo, Japan,

Department of Computational Biology, University of Tokyo, Kashiwa, Japan

Abstract. We present a partial parser called TeLePaPa (TextLens Par-

tial Parser) to identify subjects and predicate verbs (SPVs) in a sentence

of abstracts of MEDLINE citations. The performance of TeLePaPa is the

precision of 96.7% and 97.1% for the SPV detection, respectively, and

the recall of 91.3% and 94.9%, respectively. We found that there was a

similarity in the distribution of the pairs of SPV over diﬀerent research

topics in the domain. In addition, we found that the power law holds

for the relationship of the number of citations uncovered by SPV pairs

and its rank. That is, only a half of the pairs covered about 90% of all

the citations. This fact enables us to eﬃciently scan the huge amount of

biomedical literature.

1 Introduction

Thanks to the rapid development of the information technology, a vast amount

of data can be inexpensively stored and transfered as an electro-magnetic form.

The ﬁeld of biomedicine is not an exception. The researchers can obtain several

kinds of valuable information on the Internet for free, such as literature (by

PubMed

), gene (by LocusLink

), and protein (by SwissProt

In addition, the evolution of the biomedical research technology has enabled

the researchers to get lots of data all at once (i.e., Microarray). As a result, those

cases have happened more than before which not only the genes a researcher

knows well, but also those unfamiliar to him or her relate to the studying phe-

nomenon. In this situation, he or she usually begins to survey those unfamiliar

genes by exploring research papers which discuss them since essential knowledge

of biomedicine is still stored in literature. Major starting point to ﬁnd those pa-

pers is to conduct a PubMed search, where a user inputs some keywords which

he or she thinks most relevant to the concept in his or her mind. However, since

a PubMed search retrieves related literature from MEDLINE database which

stores more than 12 million citation data, the result is often too many for a

ttp://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=PubMed

http://www.ncbi.nlm.nih.gov/LocusLink/

http://www.ebi.ac.uk/swissprot/index.html

Yamamoto Y. and Takagi T. (2004).

A Practical Partial Parser for Biomedical Literature Summarization.

In Proceedings of the 1st International Workshop on Natural Language Understanding and Cognitive Science, pages 75-85

DOI: 10.5220/0002678700750085

 SciTePress

single person to check all of listed papers in a practical period of time. The cost

of learning a new gene is high and all the more if the amount of data to be

surveyed is larger.

One approach to solve this issue is to provide users with a more compre-

hensive literature retrieval system such as ExploreMed [1]. With ExploreMed, a

user can narrow down papers by interacting with the system so as to ﬁnd only

his or her interesting ones. Another approach, that is our goal, is to retrieve re-

lated literature and summarize it for a user to reduce the learning cost. We are

working on developing such a system called TextLens. This system includes text

retrieval and multi-document summarization. To summarize free text appeared

in biomedical literature, we assume to use several natural language processing

(NLP) approaches. While quite a few eﬀorts have been made for text summa-

rization [2, 3, 4, etc...], few has been done in biomedical domain, especially for

focusing on titles and abstracts (we call the title and the abstract of a paper

“a citation data” hereafter unless otherwise mentioned.). Our short-time goal is

to summarize text appeared in the a citation data obtained from MEDLINE.

This goal is to narrow down our ﬁnal goal to a single document summarization

system which only focuses on a citation data.

As for the reason why we use only citation data, we assume that they are

compiled by the authors, and that the essence of the paper, that is what we want

to extract, should be included there. Besides, to use full papers for text processing

is diﬃcult in this domain [5]. Obtaining full papers in a form for text processing

is not as easy as getting a citation data from MEDLINE. We need to get a license

from each publisher but some do not permit. Even though we can obtain full

papers, they are usually written in HTML (Hyper Text Mark-up Language) and

have not been standardized yet. For these reasons, we assumed that extracting

relevant sentences from the citation data was enough to get a main idea of a

paper and decided to use MEDLINE for our system. A relevant sentence here

means a sentence which contains information important to a biological researcher

who is seeking the contribution made by the authors of a paper.

In this paper, we introduce a partial parser called TeLePaPa (TextLens Par-

tial Parser) to ﬁnd a relevant sentence. An input of TeLePaPa is a Part-Of-Speech

(POS) tagged sentence output by the parser FDGLite [6]

, and its output is an

annotated sentence indicating the portions of its subjects and predicate verbs

(SPVs)

. To cope with long and complex collocations often seen in biomedi-

cal literature, we took an approach of chunking them before passing them to

FDGLite. The chunking method is based on Barrier word method [7]. After get-

ting a FDGLite output, TeLePaPa looks up SPVs. Our approach is a rule based

and deterministic method which iterates a replacement of a POS tag of each

word or chunk (term) with a letter denoting its syntactic characteristics such as

FDGLite is the reduced-functionality version of FDG, a full parser which outputs a

result of its deep analysis of a sentence. We use FDG to compare with TeLePaPa.

Some sentences (e.g., compound sentence) have multiple subjects and/or predicate

verbs.

noun phrase or preposition

. The iteration ends when no more rules to replace

can be applied. Following the iteration, another set of rules are applied.

We applied TeLePaPa to two corpora of citation data whose research topics

were mainly on MAP-kinase (MAPK, a protein) and Aquaporin (AQP, a human

UV-regulated gene), respectively. The result of TeLePaPa showed its precision

of 96.7% and 97.1% for the SPV detection, respectively, and the recall of 91.3%

and 94.9%, respectively. We compared results of TeLePaPa with those of FDG

and found that TeLePaPa outperformed FDG in the biomedical domain. In

addition, we evaluated the eﬀectiveness of our way of chunking by comparing

the processing time with that of FDG and Charniak parser and conﬁrmed it. An

interesting ﬁnding of our experiments is that the power-law (similar to Zipf’s

law [8]) is still eﬀective to the relationship of the numb er of citation data not

covered by SPV pairs and its rank. While a couple of approaches to ﬁnd a

relevant sentence can be thought such as using a statistical information of terms’

appearances or positions of sentences in an abstract, using the feature is an

eﬃcient way to do it.

In the following sections, we will describe our motivation, the method and

its relating studies, the result, some discussions of our study, and a conclusion.

2 Motivation

Many existing parsers such as FDG, ENGCG [9], and Charniak parser [10] are

focusing on newswire or published books. Text in such media is usually written

by professional writers, and therefore there is rarely a grammatical or syntactic

inappropriateness or any writing oﬀ the standard way.

Biomedical papers, on the other hand, are usually written by researchers

who are not professional in writing or have limited English proﬁciency as me,

and therefore text in those papers is not necessarily well written in terms of

grammatical aspects. Consequently, in order for a computer to appropriately

process biomedical papers, it should be kept in mind that not only the amount

of them is large, but also linguistic “quality” or writing style is diﬀerent from

paper to paper. This means it is more challenging to process such text.

Furthermore, biomedical literature has some characteristics diﬀerent from the

other text such as newswire (see Table 1). First, the average number of words in

a sentence is more than that of newswire. Our preliminary experiment showed

that the average number of words in the citation data was 14.4, which was 0.7

more than that of Penn Tree Bank (PTB) corpus [11], 13.7.

Second, more long collocations appear in biomedical literature. This is be-

cause many gene names, protein names, or otherwise any chemical substance

name used in the literature consist of multiple words such as JNK interacting

protein 1 or insulin-like growth factor. Our preliminary experiment showed that

the average number of words of collocations in biomedical literature was 2.9,

which was 0.4 more than that in newswire, 2.5.

Since these syntactically representing letters are really roughly assigned, they are

not necessarily grammatically correct.

Table 1. Comparison of word distributions (average). We calculated these numbers

on 56,899 sentences of citation data randomly retrieved from MEDLINE, and 52,731

sentences of PTB corpus. Collocations here means those ngrams appear more than

three times in a million ngrams (window size = 5).

# of words/stc.

# of words/colloc.

interval of longer CCs

# of colloc./stc.

MEDLINE

14.4

2.9

11.6

4.6

PTB

13.7

2.5

18.9

1.9

Table 2. Comparison of full parsers. The results of FDG above show the average

numbers of those results evaluated on subjects, objects, and predicate verbs.

FDG

Charniak Parser

Precision

93.3

96.3

93.7

91.1 (ave.)

Recall

86.0

92.0

89.0

Target

broadcast

literature

newswire

Third, more coordinate conjunctions (CCs) which have at least three coor-

dinate terms appear in biomedical literature. This means that expressions such

as “A, B, and C” or “A, B, or C” appear more frequently. Our preliminary

experiment showed that the average interval of the appearance of those CCs in

biomedical literature was 11.6 sentences, which was 7.3 less (i.e., more frequent)

than that in newswire, 18.9.

Another issue is due to existing full parsers. While a parser having an ability

to fully and correctly parse biomedical literature in a practical period of time is

ideal, to develop such a parser is quite diﬃcult. Even though a couple of existing

full parsers perform well for newswire or general books (see Table 2), they per-

form poorly when parsing biomedical literature. Our preliminary experiment on

the full parser FDG showed its accuracy of about slightly more than ﬁfty percent

(54.2%) for 192 randomly retrieved MEDLINE citation data

. In addition, full

parsing takes time to take care of the grammatical and syntactical ambiguity.

In this situation, the motivation of our developing a shallow parser is based

on our assumption that a SPV plays a vital role in a sentence for us to catch

an idea of it. Accordingly, identifying those two elements is a ﬁrst step to get

an idea the authors of a paper want to express. Therefore, while extracting an

entire and precise idea of a sentence from it by deep parsing is ideal if possible,

considering the current research environment, to identify only the two elements

still makes sense in terms of the pragmatic reasons, speed and accuracy. One

way of using TeLePaPa we assume is to search a trigger to extract a relevant

sentence from a vast number of biomedical papers.

Concerning related works, few eﬀorts have been done so far concerning rel-

evant sentence extraction from a citation data of a biomedical paper. As for

We evaluated it by checking if it correctly added a dependency function tag (deﬁned

in the user’s manual of FDG) to each word. The result was that 104 sentences were

correctly parsed out of 192.

Fig. 1. The system diagram

keyword extraction from a citation data of scientiﬁc papers, Hulth made an ef-

fort [12]. It relates to our study in that he claims the citation data is important

to catch a main idea of it.

Another related eﬀort is NLS Java Repository Project

, a project of National

Library of Medicine (NLM). They have developed a parser called PhraseX to ex-

tract noun phrases [13]. While correctly extracting noun phrases from biomedical

literature is absolutely challenging and important, TeLePaPa focuses on another

task, that is, to extract SPVs.

Pustejovsky et al. [14] developed a parser to identify and extract biomolecular

relations from biomedical literature. On “inhibition” relations, the parser showed

the precision of 90% and the recall of 57%. While this result demonstrates its p o-

tential for extracting another biomolecular relations, its performance is unknown

when it is applied to general biomedical papers to extract SPVs.

3 Method

Figure 1 shows a brief system diagram of TeLePaPa. The basic way of TeLePaPa’s

processing follows the tradition of partial parsing [15, 16, 17]. It deterministically

analyzes a sentence by applying several rules and replacing some terms with its

abstract expression in a bottom-up manner, but does not recursively apply.

Preprocessing At the beginning of the entire process, text (i.e., citation data)

is taken from MEDLINE database and split into one sentence per one line. This

process is done by a tool developed internally.

After that, domain speciﬁc words are to be chunked to minimize any mis-

parsing by FDGLite due to its lack of lexical knowledge of this domain. This is

a solution to the ﬁrst and second issues discussed in the previous section. We

built a user-made lexicon consisting of those chunks. This process is based on

Barrier word method, whose idea is that frequently appearing words can be used

as delimiters for each domain speciﬁc words such as gene name, protein name,

or any technical terms usually not incorporated in a general parser’s lexicon. In

TeLePaPa, the delimiter words are all of those included in PTB corpus.

Following the chunking, POS tagging is done by FDGLite. It adds both

syntactic and morphological tags to each word

http://umlslex.nlm.nih.gov/nlsRepository/doc/userDoc/index.html

FDG adds word-form, lemma, syntactic tag, and morphological tags.

Statistical analyses of 42 cases and 32 controls carrying the codon 31 Arg

allele identified hypertension ( odds ratio , 4.33 ) and family history of cancer

( odds ratio , 2.81 ) as positive risk factors.

Statistical analyses of 42 cases and 32 controls carrying the codon 31 Arg

allele identified hypertension ( odds ratio , 4.33 ) and family history of cancer

( odds ratio , 2.81 ) as positive risk factors.

Statistical analyses / of / 42 cases / and / 32 controls carrying the cod on 31 Arg

allele / identified / hypertension / ( / odds ratio , 4.33 / ) / a

nd / family history /

of / cancer / ( / odds ratio , 2.81 / ) / as / positive risk factors / .

Statistical analyses / of / 42 cases / and / 32 controls carrying the cod

on 31 Arg

allele / identified / hypertension / ( / odds ratio , 4.33 / ) / and / family history /

of / cancer / ( / odds ratio , 2.81 / ) / as / positive risk factors / .

Noun Phrase N, Verb V, ‘of’ 0, ‘as’ 1, ...

N0NCNVNqNcNpCN0NqNcNp1N

Fig. 2. An example of the replacement process.

Parsing TeLePaPa ﬁrstly replaces terms with a letter which basically represents

their syntactic function such as noun phrase or verb phrase (see Fig. 2). This pro-

cess goes as follows: 1) make chunks by using the following syntactic constituents

as delimiters: prepositions, conjunctions, punctuations, relatives, inﬁnitives, and

verbs, and 2) replace each chunk with a syntactically representing letter. We call

a product of this process “LS” (Letterized Sentence). Since adverbs, adjectives,

or phrases of these kinds never become a constituent of a subject or predicate

verb, they are needed to be trimmed. In addition, detecting the scope of a CC is

needed to appropriately identify multiple subjects or predicate verbs as seen in

complex sentences. These special cares are a solution to the third issue discussed

in the previous section.

TeLePaPa fulﬁlls these processes by applying a series of rules to a LS. Its

method is deterministic and therefore it can reduce a cost of the process. Each

rule consists of a regular expression and the parser substitutes a portion of a LS

with another at which the expression matches. Followings are some examples of

those rules (expressed in the form of Perl regular expression):

– M[MA]+P -> MP (e.g., will[M] be[M] precisely[A] predicted[P]),

– M+A*V -> V (e.g., could[M] have[M] sharply[A] reduced[V]), or,

– q[^p]+p -> B (e.g., ([q] i.e., ﬁve times higher )[p]).

The method to recognize the scop e a CC covers follows DP (Dynamic Program-

ming) algorithm, that is, to scan letters just before and after the conjunction

and to get scores of syntactic similarities of them.

Once getting a LS, then a process of identifying a SPV begins. When con-

sidering the metho d, we assumed that all the sentences in biomedical literature

have a SPV, and that there is no sentence having a form of inversion, that is, an

inversion of the order of a SPV. Of course, this hypothesis is not true realistically,

but such a case rarely happens especially in scientiﬁc literature. Furthermore,

since our goal is to extract a key sentence from an abstract by which the authors

want to express their contribution, it is far more rare that such a sentence is

contrary to our assumption.

Consequently, the basic pro cess of this phase is to ﬁrstly ﬁnd an independent

verb (i.e., a predicate verb) and next to ﬁnd a NP located before the found verb

(a subject). To cope with a compound sentence, the parser previously recognizes

a comma used to connect multiple clauses and iterates the independent verb

search until all the clauses are scanned.

4 Result

We developed TeLePaPa on randomly extracted sentences appeared in those

papers which mentioned human UV-regulated genes. Then, we applied it to two

corpora of MEDLINE citation data: a corpus mentioning MAPK (5,312 citations

and 51,696 sentences)and a corpus mainly mentioning AQP (776 citations and

7,916 sentences). Training set and the two corpora used to evaluate TeLePaPa

are mutually excluded. The reason why we took these corpora are their functional

importance in an organism and our internal purposes. We calculated precisions

and recalls for SPV identiﬁcations for each corpus. Those values are deﬁned as

follows: 1) Precision is equal to the number of subjects/predicate verbs correctly

identiﬁed divided by the number of those which were actually identiﬁed, and

2) Recall is equal to the number of subjects/predicate verbs correctly identiﬁed

divided by the total number of those which should be identiﬁed. Since the number

of sentences were too many for a single person to check all of them, we assessed

the TeLePaPa’s results for the randomly taken 147 sentences of the MAPK and

210 of the AQP to evaluate its performance. In addition, we evaluated the result

of FDG to assess the eﬀectiveness of our method. The FDG parsed the same

sentences of AQP as TeLePaPa did.

Table 3 shows those results. As for the MAPK sentences, both the precision

and the recall for the predicate verb detection (97.1% and 94.9%) are better than

those of the subject detection (87.2% and 89.8%). As for the AQP sentences,

however, whereas the precision of the subject detection was better than that

of the predicate verb detection (96.7% and 94.2%), the recall of the predicate

verb detection was better than that of the subject detection (92.4% and 91.3%).

Compared to the FDG output, the precision of the TeLePaPa’s output was not

as good as it, but the recall was better, especially for the predicate verb detection

(17.5% higher). This result shows that FDG takes a very conservative approach.

To assess the processing time, we implemented TeLePaPa on a Sun Fire 15000

machine and calculated it. To parse the MAPK corpus, the total elapsed time was

15 minutes including the POS tagging. As for FDG, it took 20 minutes. Moreover,

we applied Charniak parser to the same corpus. Contrary to the TeLePaPa’s

process time, it was almost a whole day long (23 hours 57 minutes).

To observe the distributions of SPV pairs, we counted the numbers of the

citation data which a pair of a SPV covers for each corpus. Besides, to see

the pairs’ distribution in general biomedical literature, we made another corpus

consisting of randomly retrieved 52,568 MEDLINE citation data published in

2002. Table 4 shows the ranks of those pairs for each category in the order of

their frequencies. While the two corpora have diﬀerent research topics from each

other, both have similar distributions. The randomly obtained corpus also shows

the similarity.

We also investigated the relationship between the number of citation data

covered by a pair and its rank. Figure 3 shows those results. The x-axis denotes

the logarithm of the rank. The y-axis denotes the number of citation data not

yet containing any already counted pairs from the top to that rank. From this

analysis, we found that the power-law holds for the relationships. As for the

Table 3. Result of TeLePaPa and

FDG. As FDG adds each tag to each

word, we regarded its parsing as good

if a NP containing a subject noun is

appropriately identiﬁed.

Precision

Recall

Subj.

Pred.

Subj.

Pred.

MAPK

87.2

97.1

89.8

94.9

AQP

96.7

94.2

91.3

92.4

FDG (AQP)

100

98.8

83.1

74.9

Table 4. The ranks of SPV pairs.

SPV Pair

Rank

MAPK

AQP

RND.

we show

these result suggest

we investigate

we demonstrate

we examine

we report

we found

4.5

5.5

6.5

7.5

8.5

0123 456

10.2

10.4

10.6

10.8

y = -0.3669x + 7.0231

= 0.9828

(left)

Aquaporin

y = -0.3373x + 8.4704

= 0.9984

(left)

MAP-kinase

y = -0.1142x + 10.919

= 0.999

(right)

Random

Uncovered Papers (log)

Rank (log)

Fig. 3. The number of citation data

not covered by SPV pairs and the rank

of a pair.

4947

146

9679

0123456

(

log

)

(

log

)

3133

7201

9771

(

, L

)

(

ubj

igh

)

Fig. 4. The number of citation data

not covered by subject or predicate

verb and their rank for “MAPK”.

MAPK pairs, the following relationship holds: log(F ) = 0.3373 log(R)+8.4704,

where F denotes the number of not covered citation data and R denotes a rank

at that number, respectively. Its correlation coeﬃcient is 0.9984, and therefore

those two numbers are highly related with each other. Let N and p

be the total

number of the citation data and a set of citation data containing the pair at

its rank of i, respectively, and S means the number of the elements the set

S contains. Since F = N

i=1

, the equation above can be expressed as

i=1

= N

, where N = 5312, C = 4771, and α = 0.3373, respectively.

As for AQP, N = 776, C = 1122, and α = 0.3669, respectively. In case of

the random corpus, the same power law still holds (the correlation coeﬃcient is

0.999), and N = 52568, C = 55215, and α = 0.1142, resp ectively.

To see whether or not the same characteristic can be found in another func-

tional constituent, we investigated distributions of subjects and predicate verbs,

respectively (Fig. 4). The right and left y-axes denote the number of citation data

not covered by subjects and predicate verbs, respectively. Those correlation co-

eﬃcients are 0.9771 and 0.9679, and therefore both of the two relationships do

not follow the power-law as much as the case of the SPV pairs.

5 Discussion

The demand on a precise and practical text processing in biomedical domain is

high. However, due to the nature of the domain, one of the fundamental functions

for it, named entity task has not answered that demand yet [5]. Nomenclature

issue has always been a great worry in this community [18, 19, 20, etc...]. In this

work, we took another approach toward the text processing, that is, rather than

tackling the named entity task at ﬁrst, grappling the basic sentence structure

detection, which has the same goal.

The drawbacks of TeLePaPa is mainly due to two issues: mal-detection of

complex sentence structures and inappropriate output of FDGLite. The former

issue can be observed when the next sentence as an example is parsed.

Our data shows that in contrast with BC2, DD1 is an inducer of X activation

measured by cell aggregation, chemiluminescence, or release of A4 serotonin.

It is a complex sentence having a that-clause, and that that-clause has several

commas; The ﬁrst two of them are used to separate the adverbial clause from the

rest, and the others are to connect each constituent of the CC. While appropriate

recognition of the function of a comma at that point is quite diﬃcult, the way of

using the second comma causes more mis-parsing currently. To ﬁnd the scope of

a that-clause is more diﬃcult if there is a comma after the “that”. Our current

parsing method of using only syntactic information has a limitation, since there

is a case that syntactic expressions or LSes of two diﬀerent sentences having

a that-clause are exactly the same, but that the scopes of the that-clauses are

diﬀerent from each other. A solution to cope with this issue without deteriorating

the simplicity should be considered.

As for the other issue, inappropriate FDGLite output, it is because there are

some cases for FDGLite to mis-annotate POS tags in biomedical literature. For

example, the word “assay” is rarely used as a verb in the domain, but FDGLite

tends to regard it as a verb. One way to compensate this problem is to make

another custom lexicon to tell FDGLite the distribution of correct POS tags for

each word frequently appeared in the domain. However, the current version of

FDGLite cannot accept that kind of lexicon.

Despite these issues, TeLePaPa achieved an acceptable performance by em-

ploying a simple approach to tackle rather complex and idiosyncratic sentences in

biomedical literature. The approach includes the in-advance chunking of domain

speciﬁc words and the parsing by the regular-expression based deterministic re-

placement. Currently, TeLePaPa only detects SPVs of a sentence, but these two

syntactic elements are fundamental to capturing the idea of the sentence. Even

though there is a case of needing deep analyses, in-advance sentence retrieval by

TeLePaPa makes them more eﬃcient by reducing the total number of sentences

to be deeply analyzed. Our discovery of the power-law in the distribution of SPV

pairs enables it. As for the case of MAPK citation data, the top 500 pairs appear

in about ninety percent (89.1%) of all the 5,312 citation data. Considering the

total number of the pairs appeared in at least three citation data is more than

one thousand (1,024), only a half of them cover almost all the citation data, and

those most frequently used pairs seem to be triggers to extract a main idea from

an abstract such as these result suggest or we conclude.

In addition, the distributions of the top pairs have similarities across the two

corpora of citation data which have diﬀerent research topics from each other.

The corpus of the randomly retrieved citation data also shows the similarity,

suggesting that the characteristic can be seen irrespective of research area in

biomedicine. Furthermore, an analysis of a long, complex sentence can be eased

by splitting it into two parts, before and after the predicate verb since in English

no modiﬁcation happens between words beyond the predicate verb. For these

reasons, it is expected to encourage better analyses of sentences in the domain

and to lead to the performance improvement of information extraction from the

enormous amount of biomedical literature.

6 Conclusion

By using the series of rules described as regular expressions to deterministically

parse a sentence, TeLePaPa accomplished fast parsing to identify SPVs without

loosing the p erformance. The most signiﬁcant part of the contribution to the

performance improvement as compared to FDG is to previously chunk domain

speciﬁc words particularly used in biomedicine and not covered by general lex-

icons. As a result of taking the approach, TeLePaPa achieved the precision of

96.7% and 97.1% for SPV detection, respectively. It also achieved the recall of

91.3% and 94.9% for the SPV detection, respectively. Subject and predicate verb

detection helps extract a most relevant sentence from an abstract of a citation

data. In addition, we found that the power-law holds for the distribution of the

number of citation data not covered by SPV pairs sorted by the order of the fre-

quency. Our work contributes to more eﬃcient text processing for the biomedical

literature.

Acknowledgments

We thank to Hiroko Ao, Kanebo Ltd. for providing us with an excellent sentence

splitter called JASMINE. We also thank to Asako Koike for reviewing this paper

and giving us invaluable comments. This work is partly supported by the grant

Grant-in aid for scientiﬁc research on priority areas (c) Genome Information

Science from the Ministry of Education, Culture, Sports, and Technology, Japan.

References

[1] Perez-Iratxeta, C., Keer, H.S., Bork, P., Andrade, M.A.: computing fuzzy associ-

ations for the analysis of biological literature. Biotechniques 32 (2002) 1380–1385

[2] Barzilay, R., Elhadad, N., McKeown, K.R.: Inferring strategies for sentence or-

dering in multidocument news summarization. Journal of Artiﬁcial Intelligence

Research 17 (2002) 35–55

[3] Teufel, S., Moens, M.: Summarizing scientiﬁc articles: Experiments with relevance

and rhetorical status. Computational Linguistics 28 (2002) 409–446

[4] Teufel, S., Moens, M.: What’s yours and what’s mine: Determining intellectual

attribution in scientiﬁc text. In: Proceedings of Joint SIGDAT Conference on

Empirical Methods in NLP and Very Large Corpora. (2000)

[5] Dickman, S.: Tough mining, the challenges of searching the scientiﬁc literature.

PLoS Biology 1(2) (2003) 144–147

[6] Tapanainen, P., J¨arvinen, T.: A non-projective dependency parser. In: Proceed-

ings of the 5th Conference on Applied Natural Language Pro cessing. (1997)

[7] Tersmette, K.W.F., Scott, A.F., Moore, G.W., Matheson, N.W., Miller, R.E.:

Barrier word method for detecting molecular biology multiple word terms. In:

Pro ceedings of the 12th Annual Symposium on Computer Applications in Medical

Care. (1988) 207–211

[8] Zipf, G.K.: Human Behavior and The Principle of Least Eﬀort. An Introduction

to Human Ecology. Addison-Wesley Press., Reading, MA (1949)

[9] J¨arvinen, T.: Annotating 200 million words: The bank of english project. In:

Pro ceedings of Coling-94, Vol. I, Kyoto, Japan (1994) 565–568

[10] Charniak, E.: A maximum-entropy-inspired parser. In: Proceedings of NAACL-

2000. (2000)

[11] Marcus, M.P., Santorini, B., Marcinkiewicz, M.A.: Building a large annotated

corpus of english: the penn treebank. Computational Linguistics 19(2) (1994)

313–330

[12] Hulth, A.: Improved automatic keyword extraction given more linguistic knowl-

edge. In: Proceedings of the 2003 Conference on Empirical Methods in Natural

Language Processing. (2003) 216–223

[13] Srinivasan, S., Rindﬂesch, T.C., Hole, W.T., Aronson, A.R., Mork, J.G.: Finding

umls metathesaurus concepts in medline. In: Proceedings of AMIA Symposium.

(2002) 727–731

[14] Pustejovsky, J., Castaﬁo, J., Zhang, J., Kotecki, M., Cochran, B.: Robust rela-

tional parsing over biomedical literature: Extracting inhibit relations. In: Pro-

ceedings of 2002 the Paciﬁc Symposium on Biocomputing. (2002) 362–373

[15] Hindle, D.: Deterministic parsing of syntactic non-ﬂuencies. In: Proceedings of

the 21st Annual Meeting of the Association for Computational Linguistics. (1983)

123–128

[16] McDonald, D.D.: Robust partial parsing through incremental, multi-algorithm

pro cessing. In Paul, S.J., ed.: Text-Based Intelligent Systems. Lawrence Erlbaum

Asso c (1992) 83–99

[17] Weischedel, R., Meteer, M., Schwartz, R., Ramshaw, L., Palmucci, J.: Coping

with ambiguity and unknown words through probabilistic models. Computational

Linguistics 19(2) (1993) 359–382

[18] Andrade, M.A., Valencia, A.: Automatic annotation for biological sequences by ex-

traction of keywords from medline abstracts. development of a prototype system.

In: Proceedings of International Conference on Intelligent System for Molecular

Biology. (1997) 25–32

[19] Ohta, Y., Yamamoto, Y., Okazaki, T., Uchiyama, I., Takagi, T.: Automatic con-

struction of knowledge base from biological papers. In: Proceedings of Interna-

tional Conference on Intelligent System for Molecular Biology. (1997) 218–225

[20] Fukuda, K., Tsunoda, T., Tamura, A., Takagi, T.: Toward information extraction:

Identifying protein names from biological pap ers. In: Proceedings of 1998 the

Paciﬁc Symposium on Biocomputing. (1998) 707–718