The intonational Structuring of Discourse

We propose a mapping between prosodic phenomena and semantico-pragmatic effects based upon the hypothesis tha t intonation conveys information about the intent ional as well as the attentional s t ructure of discourse. In part icular, we discuss how variations in pitch range and choice of accent and tune can help to convey such information as: discourse segmentat ion and topic structure, appropriate choice of referent, the dist inction between 'given' and 'new' information, conceptual contrast or parallelism between mentioned items, and subordinat ion relationships between propositions salient in the discourse. Our goals for this research are practical as well as theoretical. In part icular, we are invest igat ing the problem of intonat ional assignment in synthetic speech. 1. I n t r o d u c t i o n The role of prosody in discourse has been generally acknowledged but l i t t le understood. Linguistic pragmaticis ts have noted tha t types of i n f o r m a t i o n s t a t u s (such as g l v e n / n e w , t o p l c / c o m m e n t , f o c u s / p r e s u p p o s i t i o n ) can be intonationally 'marked ' [1,2,3,4], t ha t r e f e r e n c e r e s o l u t i o n may depend critically on intonat ion [5, 6], t ha t intonat ion can be used to disambiguate among potent ial ly ambiguous ut terances [7,8], and tha t i n d i r e c t s p e e c h ac t s may be signalled by intonat ional means [9,10,11]. Conversational analysis of natural ly occurring da ta has found tha t speakers may signal t o p i c sh i f t , d i g r e s s i o n , and i n t e r r u p t i o n , as well as t u r n t a k i n g , intonationally [12, 13, 141 . And the fact tha t intonat ional contours contribute in some way to ut terance in terpreta t ion is itself unexceptionable [8]. To date, however, identification of the prosodic phenomena involved -and the proper mapping between thcse phenomena and their semantico-pragmatic effects -has been largely intui t ive, and the intonat ional phenomena involved have not been precisely described. Here, we describe how certain of the resources of the intonational system are employed in discourse. In part icular, we discuss how speakers ' choice of p i t c h r a n g e , a c c e n t , and t u n e contribute to the i n t e n t i o n a l and a t t e n t l o n a l s t ructur ing of discourse -the way speakers communicate the relationships among their discourse goals and the relative salience of entities, a t t r ibutes , and relationships mentioned in the discourse) Our findings emerge from an intensive s tudy of a simple example of speech synthesis: the script of a computer-aided instruct ion system, TNT (Tutor 'n ' Trainer) [16], which employs synthet ic speech to tu tor computer novices in the text editor vi. Using the Text to Speech system (TTS) i17], we have been able, by systematic variat ion of pitch 1. Grosz and Sidaer [15] propose a tripartite view of discourse structure: a llngnlstlc structure, which is the text/speech itself; an attentlonal structure, including information about the relative salience of objects, properties, relations, and intentions at a given point in the discourse; and an Intentional structure, which relates dlscourse segment purposes (those purposes whose recognition is essential to a segment achieving its intended effect) to one another. range and by a principled choice of accent and tune, to highlight the s t ructure of the tutor ia l text and thus to enhance its coherence. While most studies of how intonat ion is used in discourse, have been based solely on examination of intonat ional contours found in a na tura l corpus, we have found tha t intonat ion synthesis provides a unique oppor tuni ty to manipulate the dimensions of variat ion orthogonally. Thus we can pinpoint factors crucial for a given effect and evaluate various pa t terns for a given ut terance and context. 2. T h e D o m a i n TNT was designed to teach computer-naive subjects vi, a simple UNIX screen-oriented text editor. The tu tor ia l portion provides a brief introduct ion to word processing, to general features of vi, and to the tu tor ' s help facilities; the tu tor then guides subjects through a series of learning tasks of graduated difficulty. While the overall task s t ructure is implici t in the tu tor ia l text, the subject can influence the course of the interact ion via h is /her manipulation of a set of 'helper ' keys; these keys provide hints (HINT) and reminders (REMIND) as well as the option of s ta r t ing a task over again (DO OVER) or suspending the tutor ia l temporari ly (HOLD). The fact t ha t TNT is explicitly task-oriented, 2 makes it a good test-bed for our purposes. An appropriate segmentat ion of the text, and a notion of the purpose of each segment and the hierarchical relat ionships among segments, can be independently determined from the task at hand. Also, certain characterist ics of the text presented a par t icular ly interest ing challenge for our study. First , the script contains l i t t le pronominal reference and very few so-called clue words words and phrases such as now, next, returning to, but, and on the other hand, which can identify discourse segment boundaries and relationships among segments, signal in terrupt ions and digressions, and so on [19,20]. Both of these phenomena (together with intonation) have been identified as impor tan t s trategies for communicat ing discourse structure [15,18,19]. Their vir tual absence from the text presents a convenient oppor tuni ty for test ing the power of intonat ion to s t ructure a discourse. Second, while we were not able to isolate points in the text where subjects had special difficulties, we did informally observe certain general problems with t u r n t a k i n g 3 in the tu tor -specifically, it was not always clear when the tu tor ' s turn was over which we addressed in our synthesis of the text. 3. FO S y n t h e s i s To synthesize the fundamental frequency (f0) contours for the TNT script, we used the intonat ion synthesis program 2. That is, the tutorial is organized around a series of data processing tasks, which the subject is guided through. See [18] for discussion of the characteristics of task-oriented domain discourse. 3. The process by which speakers signal that they have (temporarily) finished speaking and by which hearers interpret such signals [21].