
(Brennan, S. E. (2000). Processes that shape conversation and their implications for computational linguistics. Proceedings, 38th Annual Meeting of the Association for Computational Linguistics. Hong Kong: ACL.) Copyright © 2000, Association for Computational Linguistics. All rights reserved. Processes that Shape Conversation and their Implications for Computational Linguistics Susan E. Brennan Department of Psychology State University of New York Stony Brook, NY, US 11794-2500 [email protected] Linguists and computational linguists who Abstract formerly used made-up sentences are now using naturally- and experimentally-generated corpora Experimental studies of interactive language use on which to base and test their theories. One of have shed light on the cognitive and the most exciting developments since the early interpersonal processes that shape conversation; 1990s has been the focus on corpus data. corpora are the emergent products of these Organized efforts such as LDC and ELRA have processes. I will survey studies that focus on assembled large and varied corpora of speech under-modelled aspects of interactive language and text, making them widely available to use, including the processing of spontaneous researchers and creators of natural language and speech and disfluencies; metalinguistic displays speech recognition systems. Finally, Internet such as hedges; interactive processes that affect usage has generated huge corpora of interactive choices of referring expressions; and how spontaneous text or "visible conversations" that communication media shape conversations. The little resemble edited texts. findings suggest some agendas for Of course, ethnographers and computational linguistics. sociolinguists who practice conversation analysis (e.g., Sacks, Schegloff, & Jefferson, Introduction 1974; Goodwin, 1981) have known for a long Language is shaped not only by time that spontaneous interaction is interesting grammar, but also by the cognitive processing of in its own right, and that although conversation speakers and addressees, and by the medium in seems messy at first glance, it is actually which it is used. These forces have, until orderly. Conversation analysts have recently, received little attention, having been demonstrated that speakers coordinate with each originally consigned to "performance" by other such feats as achieving a joint focus of Chomsky, and considered to be of secondary attention, producing closely timed turn importance by many others. But as anyone who exchanges, and finishing each another's has listened to a tape of herself lecturing surely utterances. These demonstrations have been knows, spoken language is formally quite compelling enough to inspire researchers from different from written language. And as those psychology, linguistics, computer science, and who have transcribed conversation are human-computer interaction to turn their excruciatingly aware, interactive, spontaneous attention to naturalistic language data. speech is especially messy and disfluent. This But it is important to keep in mind that a fact is rarely acknowledged by psychological corpus is, after all, only an artifactÑa product theories of comprehension and production that emerges from the processes that occur (although see Brennan & Schober, in press; between and within speakers and addressees. Clark, 1994, 1997; Fox Tree, 1995). In fact, Researchers who analyze the textual records of experimental psycholinguists still make up most conversation are only overhearers, and there is of their materials, so that much of what we know ample evidence that overhearers experience a about sentence processing is based on a conversation quite differently from addressees sanitized, ideal form of language that no one and from side participants (Schober & Clark, actually speaks. 1989; Wilkes-Gibbs & Clark, 1992). With a But the field of computational corpus alone, there is no independent evidence linguistics has taken an interesting turn: of what people actually intend or understand at different points in a conversation, or why they interruptions in fluent speech. The director make the choices they do. Conversation restarts her first turn twice and her second turn experiments that provide partners with a task to once. She delivers a description in a series of do have much to offer, such as independent installments, with backchannels from the measures of communicative success as well as matcher to confirm them. She seasons her evidence of precisely when one partner is speech with fillers like uh, pauses occasionally, confused or has reached a hypothesis about the and displays her commitment (or lack thereof) to otherÕs beliefs or intentions. Task-oriented what she is saying with displays like ah boy this corpora in combination with information about one ah boy and I don't know what that is. Even how they were generated are important for though she is the one who knows what the target discourse studies. picture is, it is the matcher who ends up We still don't know nearly enough about proposing the description that they both end up the cognitive and interpersonal processes that ratifying: like a monk praying or something. underlie spontaneous language useÑhow Once the director has ratified this proposal, they speaking and listening are coordinated between have succeeded in establishing a conceptual pact individuals as well as within the mind of (see Brennan & Clark, 1996). En route, both someone who is switching speaking and partners hedged their descriptions liberally, listening roles in rapid succession. Hence, marking them as provisional, pending evidence determining what information needs to be of acceptance from the other. This example is represented moment by moment in a dialog typical; in fact, 24 pairs of partners who model, as well as how and when it should be discussed this object ended up synthesizing updated and used, is still an open frontier. In this nearly 24 different but mutually agreed-upon paper I start with an example and identify some perspectives. Finally, the disfluencies, hedges, distinctive features of spoken language and turns would have been distributed quite interchanges. Then I describe several differently if this conversation had been experiments aimed at understanding the conducted over a different mediumÑthrough processes that generate them. I conclude by instant messaging, or if the partners had had proposing some desiderata for a dialog model. visual contact. Next I will consider the proceses that underlie these aspects of interactive spoken Two people in search of a perspective communication. To begin, consider the following 1 Speech is disfluent, and disfluencies conversational interchange from a laboratory bear information experiment on referential communication. A The implicit assumptions of director and a matcher who could not see each psychological and computational theories that another were trying to get identical sets of ignore disfluencies must be either that people picture cards lined up in the same order. aren't disfluent, or that disfluencies make (1) D:ah boy this one ah boy processing more difficult, and so theories of all right it looks kinda like- fluent speech processing should be developed on the right top there's a square that looks before the research agenda turns to disfluent diagonal speech processing. The first assumption is M: uh huh clearly false; disfluency rates in spontaneous D: and you have sort of another like rectangle speech are estimated by Fox Tree (1995) and by shape, the- like a triangle, angled, and on the bottom it's uh Bortfeld, Leon, Bloom, Schober, and Brennan I don't know what that is, glass shaped (2000) to be about 6 disfluencies per 100 words, M: all right I think I got it not including silent pauses. The rate is lower for D: it's almost like a person kind of in a weird way speech to machines (Oviatt, 1995; Shriberg, M: yeah like like a monk praying or something 1996), due in part to utterance length; that is, D: right yeah good great disfluency rates are higher in longer utterances, M: all right I got it where planning is more difficult, and utterances (Stellmann & Brennan, 1993) addressed to machines tend to be shorter than Several things are apparent from this exchange. those addressed to people, often because First, it contains several disfluencies or dialogue interfaces are designed to take on more initiative. The average speaker may believe, when an internal monitoring loop checks the quite rightly, that machines are imperfect speech output of the formulation phase before processors, and plan their utterances to machines articulation begins, or overt repairs when a more carefully. The good news is that speakers problem is discovered after the articulation can adapt to machines; the bad news is that they phase via the speaker's external monitorÑthe do so by recruiting limited cognitive resources point at which listeners also have access to the that could otherwise be focused on the task signal (Levelt, 1989). According to itself. As for the second assumption, if the goal Nooteboom's (1980) Main Interruption Rule, is to eventually process unrestricted, natural speakers tend to halt speaking as soon as they human speech, then committing to an early and detect a problem. Production data from Levelt's exclusive focus on processing fluent utterances (1983) corpus supported this rule; speakers is risky. In humans, speech production and interrupted themselves within or right after a speech processing are done incrementally, using problem word
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages8 Page
-
File Size-