A GRAMMAR AND A PARSER FOR SPONTANEOUS SPEECH Mikio Nakano, Akira Shimazu, and Kiyoshi Kogure NTT Basic Research Laboratories 3-1 Morinosato-Wakamiya, Atsugi-shi, Kanagawa, 243-01 Japan {nakano, shiraazu, kogure}~atora.nt'c.jp ABSTRACT analysis method to deal with spontaneous speech. This paper classifies distinctive phenomena occur- The former method would fail, however, when new in- ring in Japanese spontaneous speech, and proposes formation is conveyed in the utterances; that is, when a grammar and processing techniques for handling the semantic characteristics of the dialogue topic are them. Parsers using a grammar for written sentences not known to the hearer. In such cases, even ill a cannot deal with spontaneous speech because in spon- dialogue, the syntactic constraints are nsed for un- taneous speech there are phenomena that do not occur derstanding utterances. Because methods that dis- in written sentences. A grammar based on analysis of regard syntactic constraints would not work well in transcripts of dialogues was therefore developed. It these kinds of cases, we took the latter approach. has two distinctive features: it uses short units as We analyzed more than a hundred dialogue tran- input units instead of using sentences in grammars scripts and classified the distinctive phenomena in for written sentences, and it covers utterances includ- spontaneous Japanese speech. To handle those phe- ing phrases peculiar to spontaneous speech. Since the nomena, we develop a computational model called L'n- grammar is an augmentation of a grammar for writ- semble Model (Shimazu et al., 1993b), in which syn- ten sentences, it can also be used to analyze complex tactic, semantic, and pragmatic processing modules utterances. Incorporating the grammar into the dis- and modules that do combination of some or all of tributed natural language processing model described those processing analyze the input in i)arallel and in- elsewhere enables the handling of utterances includ- dependently. Even if some of the modules are unable ing variety of phenomena peculiar to spontaneous to analyze the input, the other modules still output speech. their results. This mode] can handle various kinds of irregular expressions, such as case particle omission, 1 INTRODUCTION inversions, and fragmentary expressions. Most dialogue understanding studies have focused on the mental states, plans, and intentions of the par- We also developed Grass-.] ( GT"ammar ticipants (Cohen et al., 1990). These studies have for spontaneous speech in Japanese), which enables presumed that utterances can be analyzed syntacti- the syntactic and semantic processing modules of t~he cally and semantically and that the representation of Ensemble Model to deal with some of the phenomena the speech acts performed by those ntterances can peculiar to spontaneous speech. Since G~'ass-.] is an be obtained. Spontaneonsly spoken utterances differ augmentation of a grammar used to analyze written considerably from written sentences, however, so it is sentences (Grat-J, Gr'ammar for lexts in Japanese), not possible to analyze them syntactically and seman- Crass-Y-based parsers can be used for syntactically tically when using a grammar for written sentences. complex utterances. Spontaneous speech, a sequence of spontaneously There are two distinctive features of' G~'ass-J. One spoken utterances, can be distinguished from well- is that its focus is on the short units in spontaneous planned utterances like radio news and movie dia- speech, called utter'auce units. An utterance uniL in- logues. Mnch effort has been put into incorporating stead of a sentence as in Gral-J is used as a gram- grammatical information into speech mlderstanding matical category and is taken as the start symbol. A (e.g., Hayes et el. (1986), Young et al. (1989), Okada Grass-J-based parser takes an utterance unit as in- (1991)), but because this work has focused on well- put and outputs the representation of the speech act planned utterances, spontaneously spoken utterances (illoeutionary act) performed by the unit. The other have received little attention. This has partly been distinctive feature is a focus on expressions peculiar due to the lack of a grammar and processing technique to spontaneous speech, and here we explain how to that can be applied to spontaneous speech. Conse- augment (h'at-J so that it can handle them. Pre- quently, to attain an understanding of dialogues it is vious studies of spontaneous speech analysis have fo- necessary to develop a way to analyze spontaneous cused mainly on repairs and ellipses (Bear et el., 1992; speech syntactically and semantically. l,anger, 1990; Nakatani & Hirschberg, 1993; Otsuka There are two approaches to developing this kind ~; Okada, 1992), rather than expressions peculiar to of analysis method: one is to develop a grammar spontaneous speech. and analysis method for spontaneous speech that do not depend on syntactic constraints as much as the This paper first describes Grat-J, and then classi- conventional methods for written sentences do (Den, ties distinctive phenomena in Japanese spontaneous 1993), and the other is to augment the grammar used speech. It then describes Grass-Y and presents sev- for written sentences and modify the conventional eral analysis examples. 1014 1. Subcategorization rule pus vet b 1 Rule for NP (with particle) -VP constructions. he~d [ infl sentence-final } M~CH hea.d llOlIII (M head) = (FI head) c~se g& (NGM) (14 subcat) = (M subcat) U (C) sere [index *x ] sub<:at { (M adjacent) --- nil head [IO1111 (H adjacent) = nil (:~ts(! o (A CC,) (M adjunct} = (kl adjunct} sere [ index *y ] (M lexical) -- ~,lj;t(:ent rill (M sere index) ~ (H sere index) ~djun(:t nil (M sere restric) lexical yes = (C sere restric) u (H sere restric) index *e J~ ] Symbols M, C, and tt are not names of categories but f (k,ve *e) "1 variables, or identifiers of root nodes in the graphs rep- selll [ restric { (~tgent *e *x) resenting feature structures. M, C, and H correspond [ (p~tient *e *y) to mother, complement daughter, and head daughter. The head daughter's subcat feature value is a set of feature structures. 2. Adjacency rule Fig. 2: Feature strueture for the word 'aisuru' (low.). Rule for VP-AUXV constructions, Nf x particle construe tioIlS, etc, M-+AH lions in logical ff)rm in l)avidsonian style. The se- (M head) = {H head) ina.ntic represealtation ill each lexical item eonsisls of (M subeat} -- (I] subcat} a wu'iable ealled ;m inde,: (feature, (sent index}) ;rod (U adjacent) = (A) restrictions i)laced on it, (feature (selll restric)). Every (M adjacent) :- nil time a l)hrase, structure rule is ~q)lflied, l,hese restrie (M adjunct) :: (H adjunct} tions ~tre aggregated and a logical form is synthesized. (M lexlcal} - - (M sere index} = (H sere index) For exumple, let us ~gain consider 'aisuru' (love). (M sem restric} If, in the feature structure for the phr;me 'Taro ga' = (A sem restric) U 04 sere restric) (Taro-NOM), the (sen, index) value is *p a.nd gl~,, M, A, and H correspond to mother, adjacent daughter, (sere restrie) value is {(taro *p)}, after the subc.at- and head daughter. The head daughter's adjacent fea- egorization rule is al)plid the {sere restric) v~due ill ture value is unified with the adjacent daughter's feature the resulting feature str/lcture for the phrase "['aro ga strtlcture. ais.rlC (%,'o 'oves} i~ {(~ro *x) 0ov,, *e) (ag<~t *e 3. Adjunction rule *x) (patient *e *y)}. Rule for modifier modifiee constructions. (Trat-,! cowers such fundamental Jal)~mese l)henom- M~AH ena as subcategorizal.ion, passivization, interrogatiou, (M Imad) = (H he)el) coordination= and negation, and also covers copulas, (M subcat) = (H subcat) relative clauses, and conjunctions. We developed a (H adjacent) = nil parser based on (;rat-,l by using botton>u I) eha.rt (A adjunct) = {H) pursing (Kay, 1980). Unification operations are per- (M lexical) = -- (M sere index) -- {H sere index) formed by using constraint projection, ,Ul efficient (M sere restric) method for unifying disjunctive lhature descriptions = (A sem restric) U (H sem restric) (Nakano, 1991). The l)arser is inq)lemented in Lucid M, A, and H correspond to motlmr, adjunct daughter (',ommon Lisp ver. 4.0. (modifier), and head daughter (modifiee), Tile adjunct daughter's adjunct feature value is the feature structure 3 DISTINCTIVE PHENOMENA IN for the head daughter. ,IAPANESE SPONTANEOUS SPEECtI 3.1 Classification of PhcImmena Fig. 1: Phrase structure rules in (;rat-.]. We analyzed 97 telephone dialogues (about 300,000 bytes) ~d)out using ldli!]X to pl'epare docunmnts and 2 A GRAMMAR FOil. WRITTEN 26 dialogues (about i6(),O00 bytes) obtained from SENTENCES three radio lisl;ener call-in programs (Shimctzu et al., (TrM-3, a grammar for writte.n sentences, iv a uni- 1993a). We found that a.ugmentiltg the gr~:mmlal's aud fication grammar loosely based on Japanese phrase analysis methods requires taking into acconllL &{, least, structure gr~mlnar (JI'SG) (Gunji, 1986). Of Lhe six the following six phenomena in Japanese spontaneous phrase structure rules used in Gral-J, the three related speech. to the discussion in the following sections are shOWll (1)[) expressions peculiar to Japanese spontaneous in Fig. 1 in a I)A'l'll.d] like notation (Shieber, 1986)) speech, including fillers (or hesitations). ],exica] items are. represented by feature structures, (ex.) 'etto aru ndesnkedomo ._ ' 'kono fMru tar_, and example of which is shown in Fig. 2. ...' (wel], we haw'~ thenl.., this file is...) Grat-J-based p~trsers gellerate SOlllalll, iC representa- (i)2) ll~rticlc (ease pnrtiete) omission 1lhtles for relative cirCuses ~tl,d for verb-phr~tse coordi- (ex.) 'sore w,u.ashi y'a,'imasu' (I will do it.) mttions are not showll here.