anO hess sindicO issues ni tcx[ planning. 6. Texí fermatting devices: several disceurse structure rehatiens achieve. [heir eommunica[ivc purposes ...
Uncovering the structure of discaurse: Recent developments in manolo gic text gen eration Julia LAVID Universidad Complutense de Madrid
AESTRACT One of the most urgent research tasks for the discipline of discourse analysis in 990s is an interdisciplinary discussion about the types of structure that a unified the 1 theory ol discourse should account br. To date. this scientific enterprise has only been partially undertaken by researchers in the various fields of Linguisties, Sociology. Psychology, and Artificial Intelligence. to name only a few, with the result that cfforts in one area have often been ignored or insufficiently made use of by the other. An example of this is the tare adoption of computation as a rnedium lot theoretical development by linguists. The aim of this paper is lo partially bridge Ihis gap by contributing to The development of the theory of discourse from a computational perspective. Based on recent research by the aníhor and others, Ihe papa describes the state~ofdhe-att in computational research on monologic text generation, concentraíing Oil sorne of the issues which are relevant for a future unified theory of discourse structure.
EsIudi,,s Ingleses de/a Universidad Canipl’ae,zse. 2. 139/1 55. Ed¡t. (?ompli¡tcnsc. Madrid. 1994
140
.Iidhi Lav’¿d
OU‘FU N E 1. 2. 3.
4.
1.
lntroductiori The past: tirst attempts in monologic text generation Ihe prescnt: recení cxplorations ja tea planning 3.1. Textplans 3.2. Discourse relations 3.3. Schemas 3.4. Focus shift 3.5. Sentenee planning 3.6. Tea formatting Conclusion
INTRODUCTION
The study of discourse structure has been the objcct of concern of several scientifie disciplines in the last two decades. Cognitive Psychology. Sociology, Linguistics, Artificial Intelligence ami Computational Linguistics, te name only a few, ah play a ro]e in INc dynamics of disceurse organization, and Nave ah devised their own modeis te account for the strueture of disceurse. In Ihis sense, one can say that disceurse analysis is not a unitary ficid of study in itself, bat a cross-disciphine conslantly alfected by developments in several arcas. This fact, in my view. makes the studv of disceurse an exciting task, a scientific cnterprise which is open to new and sometimes simultanecus advances in different fielús, but aÑo ene which is lcss predictable and stable than other weIl-established disciplines. Within the ficid of Natural Language Processing, the relevance and applicability of the study of disceurse began to emerge as researchers faced up lo the three central questions of Computational Linguisties. The generation question is 1mw can cnt generale cohtrent lext? and involves problems of selecting material to include and organizing it hite a wehlstructured mcssage. The parsing question is 1mw can une undersíaná lexí, aná how does knowing itt struclure Iielp? and involves preblems el recognizing segments and their organization and using segments lo help with reference reselution and other tasks. The dialoguc question is líow Cali you manage 11w building oja single discaurse structure hy severa! peopie iii conjunclion? and involves issues of negotiating turns, maintaining [he initiative, and signalling acceptance or mísunderstanding. Whereas researeh on parsing has new been pursued for between three and four decades. research en language generation goes back about two decades. la this latter subarea. computational researeh en discoarse has been pursued for not longer than a decade. and has been split into monologic and dialogic disceurse, respectively ¼ With respect te dialoguc.
¿Jace verlag uit seructure of discoarse..
141
researeh has tended te focus en “plan-based” dialogues -dialogues thai occur around ceoperative plan-based endeavours sueh as tutoring, interactive explanatien, and cehlaborative action. The nature and role of [he participants’ beliefs and intentions isa current issue of researeh (Pollack, 1986; Ceben & Levesque, 1985; Grosz and Sidner, 1990) together with [he type of plans that underlie ibis type of disceurse (Lambert and Carberry, 1991; Ramshaw, 1991>. Non-plan oriented varieties of dialogie disceurse, such as everyday conversatiens, storytelhing. cte. are net eurrently siudied. Werk en menelegie disceurse. en the centrary, has cencentrated more en language-related issues such as tea planning cohesien (Merris ami Hirst, 1991). anapher (Webber, 1991) aud style (DiMarco, 1990). Within Ibis eamp. two major approaches can be identified: [he formalist ami tbe functienalist perspeetives. BoiN approaches have been developing theories wbieh are largely complementary ami whieh, hepefully, seem [o be cenverging toward a unified model of (single- ami muIti~speaker) disceurse. In [he following sections we will concentrate on!y en recent trends in monologie tea generation whicb are relevaní for a future unified theery of disceurse siructure. ,
2.
THEI PAST: FIRST ATTEMPTS IN MONOLOGJC TEXT GENERATION
‘¡be werk of B. Grosz ami (O. Siduer (Grosz & Sidner, 1986) has been considered ene of dic more influential Iheeries of disceurse siructure within the computational arena. Combining dic twe major existing appreaches te [he s[udy of disceurse, [he formalist ami [he functionalist perspectives, [bese au[hers designed a medel thai describes a three-way parallel ana!ysis of disceurse mío ihe (strucíural) segmentalien of dic utíerances, dic (funclional) siructure of interlocutor intentions, and the atientional siate. The fermalisí argument goes as fehlews: disceurse exhibits internal structure. wbere structural segments are defined by scman[ically related materia!. Sorne of tbe more influential werk is disceurse representation theery (DRT; Kamp ¡981), nad thai of Polanyi (1988), Reichman (1985>, and Cohen (1983). While [he formalist approach coneentrates en [he development of formahisms thai capture ihe properties of disceurse segments and dic disecurse siructure itsejf, ihe functionahist argume ni goes as follows: disceurse exhibits internal structure. where [he segments are defined by cemniunicative purpese. Funetienal [beories. [herefore, have focussed en dic goals of the speaker aud [be way [bose geals are reflected in [he disceurse sirueture, ofien as interrelationships between segmenis. Sorne of INc more influential funclionalisí work is rheíorical siructure theory (RST; Mann & Thempsen, 1988). and thai el Hebbs (1979), ucd Grimes (1975))
142
Ji, lía La vid
FIGURE 1 The identification scherna u TEXT, [Mckeewn 85 ID E NTIFICATION 1. 2. 3. 4. 5.
Identificatien (class & attribiítc ¡ fuactien) jA nalegy ¡ Geas! itue ncv ¡ A [trihutive / Rcnaming~ Particular -illusira Non 1 Evidence * jAmplification ¡ Analogy 1 Attributive) tParticular-ihlusír¿nion ¡ Evidencel
+
Exaniple: Elíville (Gerrnauiy) 1) Aa importantwine vihlage of [he Rheingau regien. 2) The vineyards make wines [bat are emphaticalty el [he Rheingau sty!e. 3) with a considerable weight fer a white winc. 41 Tauhenberg, Sennenhcrg and Langenstuck are ameng vineyards el note.
Anether relevant piece of research was McKeewn’s werk. Her TEXT systcm (McKeewn. ¡985) was ene of the firsí Iext general.ers which [cok discoarse sirueture mio acecuní. lí used sehemnas (predefined representations of stereotypical paragraph structtíres) te generale short texis wbich described varjeus naval ohjeets sueh as submarincs. TEXT used feur sehemas: ldentify. Describe, Compare & Contrast. and Altribulive. An example schema is shown in Figure 1. Schemas were aÑo used in ROMPER (McCoy, 1985). SEMTE.X (Résner, 1986). TAILOR (Paris. 1987) and (Ra¡nbow. 1990), Despite [beir pepularity, [he main shertceming of sehemas is, obv¡eusly. [beir inflexibility. whieh makes compulational systems unable te replan any portion of INc ten generated and reason about it. In order te address tbis shorreeming, a merbod of dynamically assenibhing eeherent texis frem basic building blocks liad te be develeped. Afier a study invelving hundreás of paragraplis, Mann & Thompson (Mann & Thompson. ¡988), proposed that a sel of about 25 relaíions sulfices te reprcsent dic relations [bat hoid within Ihe leas Ihat normally eceur ja Enghish. ‘ibis theory, called Rhetorical Strueture Theory (RST). based itselt en [he implicit assumption [bat a paragraph is enly eoherent if ahí its paris can eventually be made te fU míe ene everarching reLajen. Mos[ relaticas centain two parts, a Nucleus ([he majen central material) and a Satellite (the ancillary, subsidiary material). Eec example. dic BACKGROUND relation is given in Figure 2. I-Iewever, Rhelorieal Siructure Theery was enly a descriptive tlieorv of tea erganization, nol a generative ene. In order lo overcorne [bis prebieni, E. Hevy eperatienalized sorne relaf loas [cern RST as plans te use [bern generatively te plan paragraphs. rather titan analytically te describe paragraph structure, [bus creating a firsí [cxl sírueturer (Hevy. 1988). This was
Uncov eptienal (Divide (5. fi. cnlity)) eptíenal (Illustrate (5. fi. cntity)) Give-Analogy (5. 1-1. cntitv))
(S ami fi stand for Speaker anO ficarer rúspectively. Describe. Define, Detall, Divide, Illustrate. anO Givc-Analogv are communicalivc intcntions)
sueh as INFORM, ASK, anO ORDER. wbich are achieved hy
3.2.
Discourve relations
One cf dic central preblenis eonfronting Olsecurse work is [he number ano inferential nature of disceurse relatiens, wbicb. as Oiscussed cavilen are respensible for handhing INc language-relaícd aspecis of a text. tic need br anO use of disceurse s[ructure rehations te ensure cehictence has been acknewlcOgcd anO widely Oiscussed by researclícrs froní severa! intehhectual fields frorn hinguisis (cg., Grimes. 1975; Ouirk & Greenbaum, 1973; Sanders et aL, 1992; Redeker. 1990: Martin, 1992). toconíputatieííai linguists (cg., Hobbs, 1978; Mann & Thonípsen, 1988); te Artificial Intelhigence vesearcbers (e.g.. Schank & Abchen. ¡977: Dabígren. 1988). Taking as a peint of departure INc rclaííons defincd in Rheterieal Strueture Tbeory (Mann & Thernpsen, 1988). which were ínter extended in [-Ievvs taxenemizatíen of a cehleetien of [he íelations prepesed by ever thirty researcbers froní varicus fictOs. ínter reorganized wdh Maler (Hevy. -
Cincovcring dic tarar-tare of disco,írse. - -
147
1990: Maier anO Hovy. 1991), anO Martin’s [axenemuza[ien of tibe conjunctive rclatiens (Martin. 1992), a recent taxenemy has been preposed te handie dic requirenienis of tihe new text pi-anner (Hevy et al.. 1992). The rcla[iens wcre divided inte [liree majer ne[werks, according te Halhiday’s me[afunc[ional sphi[: seman[ic/ideatienal, in[crpersonal. anO presen[atienal/tex[uah. ‘ihe core set of rela[ions, organized in[o a tuxenemy, are repredueed in Figure 4 As shewn iíí ihe nelworks, dic selcetien of ideational, interpersenal anO [exíuah rela[iens is net exclusive. A disceurse segment mus[ be able [o maintain thrce in[ersegment rehations simultaneously (see Lavid & Majer, 1992). .
3.3.
Sefíemas
As plan and relatien libraries grew, [he number of possible texts grews alarmingly. Se, as argued in for example (MeKeewn, 1985, Rambow. h990, Meeney eL al., 1990), ene sheuld capture Llie idiosyncratic regularities of disceurse struc[ure, which mnay depend en genre, demain, er even simply eustom, in schernas ami use 11cm as frozen plaus Ny simple schcnia ins[untiatien. Where addi[ienal s[rueturing is required -when no frezen plan exists te achieve [be communica[ivc intention- thcn disceurse s[ructure phans anO in[erscgment relatiens can be used. A sebema is a predefined represen[ation of a s[ereetypical paragraph struc[ure wbich actis as a [emplate te mandale [he content anO erder of the chauses in a paragraph, or [he erder of a series of text spans in a tex[. Tbis netion has been captured iii linguistics anO reccived different names: cg., macrostructures (Van Dijk anO Kintschi, 1983), holistie structares (Mann & Tbempson. 1988) er (Jenerie Structure Potential (Husan, 1977, 1984); in [Nc world of Artificial Intelligence anO Compu[a[ienal Linguistics, the netiens of seript (Sebank anO Abelsen, 1977) anO scherna (McKeewn, 1985), respectively, Nave a similar func[ien. Altbeugh schemas are a useful anO widcspread mechanism for [ext planning, seme unreselved issucs still remain. As ye[ no representation for scbemas captures wehl tibe underhying semaníic anO rbcterical intcrrelations of tibe partis. AÑo, U is quite diffieult te fermalize the criteria fer cen[relling [he inclusien of additional grew[h pointis [e tibe disceurse, [hat is, [be type anO erder of material te include. Tbe enly criterien wbicb has been considered, te seme ex[ent, is tihe effec[ of focus sbift en disceurse structure, and it is te this tbat we new turn.
3.4.
Focus shift
Textual ceherence isa complex plienemenen wbich cannet be captured in i[s [e[ality by ene single [ype of s[ruc[ure; it results frem [be in[erac[ion
itt/za Lavid
148
FIGURE 4 of disceurse segmení rcha[iens (N’Laier and Hovy. 1991)
A [axonemy
ElabObjccí
ElafiParí Elabora lien ElabG,encraílitv
Gircumsta ncc Semantie Sc quencc
{ {
idenl i ficali en
I{
u
Re st ate mc n
Olijeel Atíribute Objeci Funlien Sct Member 1’re cess - Sicp WNole- l>art Cien 1-5 pecilie Ab síu - Instancc Surninatv
Fi me Mean 1 nstrUineni [Sayalleí Ev cnt ScqTcrnporal SeqOrdinal
Cause ¡ Resulí
General Condition
[c~&x’o¡ (¿VR Non Vol
{
urpose
Conddlon
r Otherwise
Gempar-ative
Anal egy
Inierpersonal
{
Inteupretatien Enabiemení Antilbesis Exho rialion
Logical rehation Freseni ati onal
Presentaliena] sequence {iei
ti 1
{ {
Support Concessien Cual ¡ ficalien Conjuntion Di si u ti elien
{
Solutionhued Evidence J usti ticaíi en Melivatien
tinco ver/ng the stri,cti
