(pdf)

1997

The paper uses a simple and abstract characterization of dialogue in terms of mental state changes of dialogue participants to raise three fundamental questions for any theory of dialogue. It goes on to discuss currently popular accounts of dialogue with respect to these three questions. Next, the notion of conversational game' is revisited within a probabilistic and decision theoretic framework, and it is argued that such an interpretation is plausible both intuitively and as the basis for computational implementation. An illustrated sketch of a proposed implementation using Bayesian networks is described. Three Questions for Dialogue A simple, rather abstract description of a canonical dialogue is that it consists of a sequence of utterances with a corresponding sequence of mental states of the participants in the dialogue. Person A has a sequence of mental states SAI.. San-E1 and person B also has a sequence SB1 Sgn+1. Connecting these two sequences is a third sequence, the sequence of utterances. UA1 is produced by A in state Al, Ug2 is produced by B in B2 and so on. Furthermore, A's state SA2 and B's state SB2 are, at least partially, determined by the utterance UA1 which precedes them. The utterances change the mental states of the participants to the point where no further communication is regarded by them as necessary: the goals of the conversation, whatever they were, have been achieved as far as is possible. This is represented by the diagram in figure 1. Even this simple picture reveals that there are several large questions to be answered in order to be in a position to build a machine capable of playing the part of A or B: (i) what are mental states? (ii) how do they change? (iii) how do utterances connect with them and change them?

Travaux neuchâtelois de linguistique

Computational Linguistics, 2004

2012

Abstract On one familiar and traditional picture, linguistic communication is a matter of the expression and transmission of a proposition across a common ground, with the proposition determined as a function of its semantic value. What general properties of a system of linguistic communication indicate whether or not it can, even in principle, be modeled along these traditional lines?

Working Papers in Linguistics Lund, 1995

MARIANNE GULLBERG strong or heavy association is likely to result in the same or 'right' interpretation frequently. A number of cues will help indicate these associations. The notion of competition between different sources of information has been exploited in a model of purely linguistic sentence comprehension within the framework of the Competition Model (e.g. MacWhinney 1987; 1989). It has been used to explore variation in language behaviour, sentence processing in learner language, in bilinguals, in aphasia, etc. Cues to interpretation combine or compete in discourse and at every moment a decision has to be made as to what or who the referent is-what the message is. When interacting, cues form clusters which are related to and conditioned by other clusters and certain cues demand certain other cues for optimal interpretation. Note that the key word here is optimal. What distinguishes this from traditional feature analysis, subcategorisation frames, and the like, is the connectionist/associationist 2 view that there is no fixed outcome of the weighting of cues in conflict or in cooperation , but that the best possible interpretation at any given moment will be the result of weighting clusters for or against a certain interpretation. Interpretation is not rule-based, but probabilistic. This helps explain variation in language use and how we deal with it, why we understand anomalies or creative language like do a Napoleon for the camera (Clark & Gerrig 1983), etc. Recently it has been suggested that discourse phenomena might also be treated in terms of cues. St. John 1990 has simulated comprehension of a text using a construct called cue-constraint satisfaction, where interrelated cueclusters condition each other. At discourse level, cue-clusters can be assumed to help resolve co-reference problems, e.g. cues may cluster to indicate the likely referent, and the strongest cluster will successfully designate the referent. This is a convenient way of handling what has been referred to as 'context' or 'world knowledge' related problems. Scripts, frames and conceptual structures (Brachman 1977; Minsky 1975; Schank & Abelson 1977) are all constructs trying to deal with what Tannen calls our expectations of the world (Tannen 1993). So far, only a limited set of cues have been investigated: lexical/semantic information, word order, morphology and prosody. When introducing cue-based comprehension at discourse level, however, cues related to world knowledge will have to be introduced. What constitutes a cue in

4 Conclusion In the previous sections we showed that the ProPars system by using symbolic procedures is enabled to parse sentences of arbitrary length. In addition, symbolic procedures were used to store intermediate parsing results such that a syntax tree can be assembled after the parsing process has been completed. Both of these problems are not solved by other connectionist based systems.

Speakers have been hypothesized to organize discourse content so as to achieve communicative efficiency. Previous work has focused on indirect tests of the hypothesis that speakers aim to keep per-word entropy constant across discourses to achieve communicative efficiency (Genzel & Charniak, 2002). We present novel and more direct evidence by examining the role of topic shift in discourse planning. If speakers aim for constant per-word entropy, they should encode less unconditional per-word entropy (as estimated based on only sentence internal cues) following topic shifts, as there is less relevant context to condition on. Applying latent topic modeling to a large set of English texts, we find that speakers are indeed sensitive to the recent topic structure in the predicted way. Keywords: discourse production; topic shift; communicative efficiency

The opening remarks and first lectures assigned to any typical introductory course in linguistics (i.e., child language acquisition, syntax) often invest a sizable amount of time with the analysis of the hidden structure of language. The range of analyses can consist, from a minimum scale, of the simple observation of how language is recursive in nature (my Russian Doll note [1]), to a maximum-scale, of the entire fullsweeping presentation of Chomskyan-style syntax [2, 12]. Out of such first-course analyses of linguistic structure, various vignette discussions emerge: e.g., (i) the classic Skinner v Chomsky debate [3] (where the first-generation data of child utterances are introduced to the student: viz., Berko's 'Wugs test' [3]), (ii) morphological analyses and distinctions between derivational vs inflectional morphology [4], (iii) the development of morphosyntactic structure in child language [5,6,7,10], leading to (iv) maturational hypothesis of syntactic structure as pegged to the neuro-onsets of specific regions in the brain [3,13]. Concluding lectures often attempt to summarize actually 'what it is that makes language interesting' [8], while final 'accumulative lectures' attempt to show how such structure is widely pervasive throughout language in general [9]. In an overall sense, the Skinner v Chomsky debate as typically found in introductory lectures makes-up for a fine 'pedagogical device' in framing much of the discussion on language structure, upon which a maturational hypothesis as pegged to brain development can be easily overlapped. The material arranged here represent an array of pdf-lectures and chapter-readings on the topic of Linguistic Structure (in preparation for 'Linguistic Essays on the Topic of Structure'.

The profound use of the computer in discourse analysis must employ a theory of discourse comprehension and production with which to conduct the analysis. Models currently employed in computational linguistics have a semantic basis and are goal-directed. The basic model is an associative cognitive network. The basic inventory of concepts of the system is given in the systemic network, which is organized into paradigmatic, syntagmatic, and componential structures. Since events happen in particular places at particular times, there is also an episodic structure. The gnomonic system defines abstract concepts over episodes. According to Phillips (1975), discourse coherence must be considered on two levels, the episodic and the gnomic. A discourse which engenders episodic and/or gnomonic expectations which are not then fulfilled is incoherent. A lower limit on coherence may be defined as a discourse so ill-formed that it makes no sense even to its creator. The upper limit on coherence is set by the most powerful creative minds. Between the two limits, discourse analysis, from the point of view of the computational linguist, probably requires nothing less than a full-blown computational theory of the human mind. (JB)