Kenjl Ono, Kazuo Sumlta, Seijl Miike
Research and Development Center
Toshiba Corporation Komukai-Toshiba-cho 1,Saiwai-ku, Kawmsaki, 210, Japan
ono@isl.rdc.toshiba.co.jp 1 ABSTRACT We have developed an automatic abstr~mt generation system for Japanese expository writings based on rhetorical structure extraction. The system first extracts the rhetorical structure, the compound of the rhetorical relations between sentences, and then cuts out less important parts in the extracted structure to generate an abstract of the desired length. Evaluation of the generated abstract showed that it contains at maximum 74% of the most important sentences of the original text. The system is now utilized as a text browser for a prototypicaI interactive document retrieval system. 2 INTRODUCTION Abstract generation is, like Machine Translation, one of the ultimate goal of Natural Language Processing. IIowever, since conventional word-frequencybased abstract generation systems(e.g. \[Kuhn 58\]) are lacking in inter-sentential ordiscourse-structural analysis, they are liable to generate incoherent abstracts. On the other hand, conventional knowledge or script-based abstract generation systems(e.g. \[behnert 801, \[Fum 86\]), owe their success to the li,nitation of the domain, and cannot be applied to document with varied subjects, such ,as popular scientific magazine. To realize a domain-independent abstract generation system, a computational theory for analyzing linguistic discourse structure and its practical procedure must be established.
ltobbs developed atheory in which lie arranged three kinds of relationships between sentences from the text coherency viewpoint \[Hobbs 79\].
Grosz and Sidner proposed a theory which accounted for interactions between three notions on discourse: linguistic structure, intention, and attention \[C, rosz et al. 86\].
l,itman and Allen described a model in which a discourse structure of conversation was built by recognizing a participanUs plans \[Litman et al. 87\]. 'l'hese theories all depend on extra-linguistic knowledge, the accumulation of which presents a problem in the realization of a practical analyzer.
Cohen proposed a framework for analyzing the structure of argumentative discourse \[Cohen 87\], yet did not provkle a concrete identification procedure for 'evidence' relationships between sentences, where no linguistic clues indicate the relationships. Also, since only relationships between successive sentences were considered, the scope which the relationships cover cannot be analyzed, even if explicit connectives are detected.
Mama and Thompson proposed a linguistic structure of text describing relationships between sentences and their relative importance \[Mann et al. 87\]. llowever, no method for extracting the relationships from superficial inguistic expressions was described in their paper.
We have developed a computational rnodel of discourse for Japanese expository writings, and implemented a practical procedure for extracting discourse structure\[Sumita 92\]. In our model, discourse structure is deiined ,as the rhetorical structure, i.e., the compound of rhetorical relations between sentences in text. Abstr~t generation is realized ~s a suitaMe application of the extracted rhetorical structure. In this paper we describe briefly our discourse model and discuss the abstract generation system based on it.
Rhetorical structure represents relations between varions chunks of sentences in the body of each section. In this paper, the rhetorical structure is represented by two layers: intra-paragraph and inter-paragral)h structures. An intra-paragraph structure is a structure whose representation units are sentences, and an inter-paragraph structure is a structure whose representation units are paragraphs.
In text, various rhetorical patterns art,. used to clarify the principle of argument. Among them, co,> nectivc expressions, which state inter-sentence relationships, are the most significant. The tyl)ieal grantmatical categories of the connective expressions are connectives and sentence predicates. They can I>c divided into the thirty four categories which are exernplified in Table 1. Table h Example Of rhetorical relations
Relation Expressions serial (<SR>) dak'ara (thus) su'n{lnarizatiou 'kh.a.,aOk.,, (after all) (<su>)
negative (<NG>)
example {<EG>)
espeeial(<ES>)
re.~son !<aS>) s{ipplcment (<SP>) background (<BI>)
parallel (<PA>) exteflsion (<EX>) rei)hra~e (<RF>) direction (<DI>)
The rhetorical relation of a sentence, which is the relationship to the preceding part of the text., can be extracted in accordance with the connective expression in the sentence. For a sentence without any explicit connective exl)ressions , extension relation is set to the sentence. The relations exemplitied in Table 1 are used for representing the rhetorical structure.
Fig. 1 shows a paragral)h from an article titled "A Zero-Crossing l{ate Which Estimates the Frequency of a Speech Signal," where underlined words indicate connective xl)ressions. Although the fourth and fifth sentences are clearly the exemplification of the first three sentences, the sixth is not. Also the sixth sentence is the concluding sentence for the shikashi (I)ut) tatoeba (for example) tokuni (particuli~rly) ,mzenara (because) mochiron (of course) juurai (hitherto) mata.(and) kore wa (this is) tsumari (that is to say) k'okode wa ... wo nobeT~l (here ... is described) In tile context of discrete-time signals, zerocrossing is said to occur if successive samples have dilfereut algebraic signs. Tile rate at which zero crossings occur is a simple measure of tile frequency content of st sig,ml. This is .particularly true of narrow band signals. For example, a si,msoidal signal of frequency P0, sanll)led at a rate fs, h,'~s i'~/t"~ samples per cycle of the siue wave. Each cycle has two zero crossings so that the hmg-term average rate of zero-crossings is z = 2F0/s;;. Thus, tile average zero-crossing rate gives a reasonable way to estinmte the frequency of a sine wave. (L.lt.l(abiner and \[{.W.Schafer, Digital l','ocessing of Speech Siffmtls, Prentice-llall, 1978, p.127.) Figure 1 : Text example first five. Thus, tile rhetorical structure for this text can be represented by a binary-tree as shown in Fig. 2.This structure is also represented as follows: \[\[\[1 <EZ> 2\] <gs> \[3 <E(\]> \[4 <EX> 5\]\]\] <sa> 6\] 1: 2: 3: 4: 5: 6: Figure 2: Rhetorical structure for the text in l,'ig.1
The rhetorical structure is represented by a binary tree on the analogy of a syntactic tree of a natural language sentence. Each sub tree of the rhetorical structure forms an arg,rnentative constituent, just as each sub-tree of tile syntactic tree forms a gram,natical constituent. Also, a sub-tree of the rhetorical structure is sub-categorlzed by a relation of its parent node as well as a syntactic tree.
The current version of TECIIDOC is running on Sun Spare stations with LUCI\]) CommonLISP 1.4 and LOOM 1.41 (a port to LOOM 2.1 is underway), and a PENMAN version fl'om 199i. The user interface is based on the CommonbISP Motif interface package CI,M and the application building tool GINA \[Spenke ct al., 1992\]. Acknowledgements ~l'he success of the TECIIDOC i)rojeet depended heavily on eontril)utimls from a lltllllb(!r of student interns, in alphabetie;d order: Brigit.te Grote,, Sitll(Ira Kiibler, Itaihua Pan, .lochen Schoepl>, Alex~mdot Sigel, Ralf Wagner, and Uta We, is. ~i'hey ~dl have contributed to gl'&lltll'Lar or le×icon coverage ill one wa~y or another. Qerhard Peter has implemented TI'~CtlDOC-I, an intera(:tiw~ version giving c~tr mMntainanee ~tssist;tnce. Thorsten Liebig hats imph~mented TECtlDOC's user interface for workstatim~s using CLM and GINA, Ilartmut Peuehtmiiller has ~t(tded multimedia facilities ~md mouse-sensitive t xt mltlmt. We also have to thank the PlgNMAN ~tn(l LOOM groups ~tt USC/ISI and the KOMET project ~tt GMD Darmstadt, wire gave us inwdmd~te help. References \[Bateman, 1990\] ,h)hn A. Bateman, Upper model-
ing: A level of semantics for n;tt~lrltl l~tngu~tge
processing. In PTvcecdings of the Fifth hJter.n,-
tional Workshop on Nahu'al Lang~tagc G'eneration,
Pittslmrgh, PA., 3 - (; June 1909. \[Grote et al., 1993\] Brigitte Grote, D~etmar llSsner,
~tnd Manfred Stede. ll.epresentation lewzls in mul-
tilingual text genera.tion. In Brigitte (\]r(~te, Di-
etmar R.i~sner, Manfred Stede, and Uta Wets, edi-
tors. From l(no'wledge 1o L~t~gmtge Three l)~tpers
on MMtiling~ml tea:t Ge*teration. FAW Uhn, FAW-
T11.-93017, 1993. \[LOOM, 19911 q'h,~ LOOM l~nowledge l~.present;t-
t, ion Systellt. \])oettlllell~,~ttil)ll \]~;tcklt\[! i!~
USC/Information Sciences Institute, Marina l)el
I{ey, CA., 1991. \[Mann and Thompsm~, 1987\] Willi;tm C. Mam~ and
Sandra A. Thompson. IlhetoricM structure the-
ory: A theory of text ()rg;tnization. In L.Pohmyi,
editor, 7'he Sl*"uctttre of Discmtrse. Ablex, Nor-
wood, N.J., 1987. Also as USC/Informatim~ Sci-
ences Institute Research Report IIS-87-t90. \[\[I.i~sner and Stede, 1992;~\]
Dietmar Ri~sner \[tnd Manfred Stede. Customiz-
ink I1.ST for the automatic production of tech-
nical manuals. In R. D;tle, \]'~. Ihwy, D. l/Sslw.r,
and O. Stock, editors, Aspects of A'utomatcd Nat-
'ltral Language Generation - l)roeeeditlos of the 6tb lnter~mtio',.al WS (m Natural LaTLg'uaqe Geneva~ic, n, Lecture Notes in Artificial Intelligence 587. Springer, llerlin/l\[eidelberg, 19(.12. \[IlSsner itnd SLed.e, 1992b\] Dh, tm~u' I/Ssner ;tn(|
Manfred Stede. TEC\[ll)OC : A system fi~r the au-
t.mnatic l~roduction of multilingual technical doc-
uments. In C,, Giirz, editor, KONVENS' 92, Reihe
\[nformat.ik aktuell. Springer, l~erlin/Ihfidelherg,
19\[)2. \[Spenke t ,l., 1992\] Miehltel Spenke, Christian
\[~eilken, 'Phomas Berlage, And.'e~s Bi\[cker, ~tnd
Andreas (\]rau. UlNA lh'feve',ce Ma'n'~utl Versio'n
2. I. G~wmlm Ni~tion;d F/esea.rch Center for Con>
purer Science, Snnkt Augustin, Cb~rmany, 19(.12. that case the system cuts out terminal nodes from the last sentences, which are given the same penalty score.
If the text is written loosely, tile rhetorical structure generally contains many BothNuelevs relations (e.g., parallel(marc(and, also)), and the system cannot gradate the penalties and cannot reduce sentences smoothly.
After sentences of each paragraph are reduced, inter-paragraph structure reduction is carried out in the same way based on the relative importance ju~lgement on the inter-paragraph rhetorical structure.
If the penalty calculation mentioned above is accomplished for the rhetorical structure shown in Fig. 2, each penalty score is calculated as shown ill Fig. 3. In Fig. 3 italic numbers are the penalties the system imposed on each node of tile structure, and broken lines are the boundary between the nodes int-posed different penalty scores. The figure shows that sentence four and five have penalty score three, that sentence three has two , that sentence one and two have one, and that sentence six has no penalty score. In this ease, the system selects sentence one, two, three and six for the longest abstract, and and also could select sentence one, two and six as a shorter abstract, and also could select sentence six as a still more shorter abstract.
After the sentences to be included in tile al)stract are determined, the system alter,atcly arranges the sentences and the connectives from which the relations were extracted, and realizes the text of tile abstr~t.
The important feature of the generated abstr,'mts is that since they are composed of the rhetoriealy consistent units which consist of several sentences and form a rhetorical substructure, the abstract does not contain fragmentary sentences which can,ot be understood alone. For example, in the abstract generation mentioned above, seutence two does not al> pear solely in the abstract, but appears ahvays with sentence one. If sentence two apl)eared alone in the abstract withont sentence one, it wouhl be difficult to understand the text. 6 EVALUATION The generated abstracts were evaluated from the point of view of key sentence coverage. 30 editorial articles of"Asahi Shinbun", a Japanese newspaper, and 42 technical papers of "Toshiba Review", a journal of Toshiba Corp. which publishes short expository papers of three or four pages, were selected and three subjects judged tile key sentences and tile most important key sentence of each text. As for the cditohnport. Node right node left node both nodes q'able 2: R.elative importance of rhetorical relations Relation Type RighlNncleus LeflNvclens llothNuclcus Ij ; $ Figure 3: Penalties on relative iml)ortance for the rhetorical structure in Fig.2 rial articles, The average correspondence rates of the key sente.ce and tile most important key sentence among the subjects were 60% and 60% respectively. As for the technical )apcrs, they were 60% and 80 % resl)ectlvely.
Then tile abstracts were generated and were compared with the selected key sentences. The res,lt is shown in Table 3. As for the technical papers, tile average length ratio( abstract/original ) w;~s 24 %, and tile coverage of tl,e key sentence and the most important key sentence were 51% and 74% respectively. Whereas, ~s for the editorials, tile average length ratio( abstract/original ) was 30 %, and the coverage of the key sentence and the most important key sentence were 41% and 60% respectively.
The reason why the compression rate and the kc.y sentence coverage of the technical papers were higher than that of the editorials is considered as follows. The technical papers contains so many rhetorical expressions in general as to be expository.
ltelation serial, smnmariz~ttion, negative, ... exalnplc~ reason, especial, SUl)plernen t, l)arallcl, extension, rel)hrase , ... That is, they provide many linguistic clues and the system can extract the rhetorical structure exactly. Accordingly, the structure can be reduced further and the length of the abstract gets shorter, without omitting key sentences. On the other hand, in the editorials most of the relations between sentences are supposed to be understood semantically, and are not expressed rhetorically. Therefore, they lack linguistic clues and the system cannot extract the rhetorical structure exactly. Table 3: Key sentence coverage of the abstracts cover ratio Material total length
num. ratio key I \] mosl. sentence iml)°rtant Sell \[+etlee
editorial 30 0.3 0.,11 0.60 (Asahi Shlnbun)
tech. journal 42 0.24 0.51 0.7.1 (Toshiba Review) 7 CONCLUSION We have developed an automatic abstract generation system for Japanese expository writings based on rhetorical structure extraction.
The rhetorical structure provkles a natural order of importance among senteuces in the text, and can be used to determine which sentence should be extracted in the abstract, according to the desired length of the abstract. The rhetorical structure also provkles the rhetorical relation between the extracted sentences, and can be used to generate appropriate connectives between them.
Abstract generation b~sed on rhetorical structure extraction has four merits. First, unlike conventional word-frequency-based abstract generation systems(e.g. \[Kuhn 58\]), the geuerated abstract is consistent with the original text in that the connectives between sentences in the abstract reflect their relation in the original text. Second, once the rhetorical structure is obtained, varions lengths of generated abstracts can be generated easily. This can be done by simply repeating the reduction process until one gets the desired length of abstract. Third, unlike conventional knowledge or script-b`ased abstr,~t generation systems(e.g. \[Lehnert 80\], \[Fum 86\]), the rhetorical structure extraction does not need prepared knowledge or scripts related to the original text , aud can be used for texts of any domain , so long as they contain enongh rhetoricM expressions to be expository writings. Fourth, the generated abstract is composed of rhetoriealy consistent units which consist of several sentences and form a rhetorical substructure, so the abstract does not contain fragmentary sentences which cannot be understood alone.
The limitations of the system are mainly due to errors in the rhetorical structure analysis and the sentence-selection-type abstract generation, the evalnation of the accuracy of the rhetorical structure analysis carried out previously( \[Sumita 92\] ) showed 74%. Also, to make the length of the abstract shorter, It, is necessary to utilize an inner-sentence analysis and to realize a phrase-selection-type abstract generation b,~sed on it. The anaphora-resolution and the toplc-sul)l)leineutation must also be realized in the analysis.
The system is now utilized ,as a text browser for a prototypical interactive document retrieval system. References \[Cohen 87\] Cohen, I/. : "Analyzing the Structure of Ar-
gtnnentative Discourse", Computational Linguistics,
Vol.13, pp.ll-24, 1987. \[Fum 86\] Finn, D. : "Tailoring Importance Evalnatlon
to Reader's Goals: A Contribution to Descriptive Text
Summarization", Proc. of Coling, i117.252-259, 1986. \[Grosz et al. 86\] Grosz, B.J. and Sidner, C.L. : "Atten-
tion, Intentions and the SLruct.ore of Discourse", Com-
putationM Linguistics, Vol.12, pp.175-204, 1986. \[Ilobbs 79\] Hobbs, J.R.: "Coherence and Coreference",
Cognitive Science, Vol.3, 1979, pp.67-90. \[Kuhn 58\] Knhn, II.P. : "The Automatic Creation of Lit-
erature Abstracts", IBM Journal, Apr. 1958, pp.159-
165. \[I,elmert 80\] l,ehnert, W. : "Narrative Text Summariza-
tion", Prec. of AAAI, I111..'137-339, 1980. \[Litman et al. 87\] l,ltman, D.J. and Allen, J.F.: "A Plan Recognition Model for Subdlalogues in Conversations", Cognitive Science, Vol.ll, 1987, pp.163-200. \[Mann et al. 87\] Mann, W.C. and Thompson, S.A. :
"Rhetorical Structure Theory: A Framework for the
Analysis of Texts", USC/lnfi)rmation Science lnsti-
lute Research Report RI/-87-190, 1987. \[Sumita 92\] Sumita, K., et al. : "A Discourse Struc-
ture Analyzer for Japanese Text", Pr~. h~t. Conf.
Fifth Generation Computer Systems 1992 (FGCS'92),
pp.1133-1140, 1992.