The Many Functions of Discourse Particles: A Computational Model of Pragmatic Interpretation

We present a connectionist model for the interpretation of discourse particles in real dialogues that is based on neuronal principles of categorization (categorical perception, prototype formation, contextual interpretation). It can be shown that discourse particles operate just like other morphological and lexical items with respect to interpretation processes. The description proposed locates discourse particles in an elaborate model of communication which incorporates many different aspects of the communicative situation. We therefore also attempt to explore the content of the category discourse particle. We present a detailed analysis of the meaning assignment problem and show that 80% – 90% correctness for unseen discourse particles can be reached with the feature analysis provided. Furthermore, we show that ‘analogical transfer’ from one discourse particle to another is facilitated if prototypes are computed and used as the basis for generalization. We conclude that the interpretation processes which are a part of the human cognitive system are very similar with respect to different linguistic items. However, the analysis of discourse particles shows clearly that any explanatory theory of language needs to incorporate a theory of communication processes. Discourse Particles, Meaning Assignment, and the Communication System In a number of papers (Scheler, 1994; Scheler, 1996b), it has been argued that feature-based categorization is an effective model for morphological and lexical meaning analysis. This applies to the issue of generating a morphological category or a lexeme from a given feature representation, but it is even more effective in extracting meanings from the context of a written text or dialogue. The central idea is that linguistic elements are used in communication (whether with self or others) via reference to aconceptual level, which is closely tied to cognitive categories (event and temporal structure, spatial cognition, knowledge about objects, motion, changes etc.). In this paper, it will be argued that pragmatic meanings basically operate in the sameway: a limited number of discourse elements in a language serves to communicate a greater number of pragmatic meanings in a predictable way. This model of categorization as meaning assignment will be instantiated for discourse particles, characteristic items of spoken language discourse, such as well, yes, oh, ah, okay, uh and um. These elements fulfillmany different functionswith respect to the turn-taking system, speech management, discourse structure, and the interactive level between the communication partners. For example, they segment utterances; they indicate new topics and mark important information; they establish a harmonious atmosphere between speakers and hearers; they help taking, yielding or holding the turn, and they signal speaker-attitude (Levelt, 1983; Schiffrin, 1987; Grosz et al., 1989). Most analyses of these particles are characterized by restricted perspectives on their function, which is mirrored in a large number of terms for the phenomena under consideration, for instance segmentation marker, cue phrase, connector, interjection (Fischer, 1996b). Additionally, so far no automatic methods for assigning discourse functions to discourse particles have been proposed. Using a specific model of the communication situation (Fischer, 1996a), a method for assigning discourse functions to particles in natural dialogues will be presented which is based on neuronal categorization principles as embodied within the connectionist framework of NEUROSEM. The aims of the paper are therefore to show that feature-based categorization is an effective model for pragmatic interpretation processes; to automatically assign meanings to occurrences of discourse particles in context; to show how knowledge about the functions of one discourse particle can support another by means of generalization from prototypes. A Computational Model of Discourse Function The linguistic model developed for discourse particles was motivated by the goal of finding consistent patterns of the pragmatic function of these particles, which seem to occur almost randomly at first sight. However, our hypothesis was that if we apply a computational model that is capable of accounting for a high degree of context-dependency and multifunctionality, the different meanings of discourse particles can be computed with a considerable degree of accuracy. In the followingwe shall present briefly the linguisticmodel and the computational model applied to it. We refer the reader to (Fischer, 1996a) and (Scheler, 1997) for a fuller account of these approaches. The Linguistic Model In contrast to studies on discourse particles so far, the model proposed here is not restricted to a particular function discourse particles fulfill in spoken language discourse but it considers the whole range within a model of communication. The model proposed treats discourse particles as lexemes with an invariable cognitive content which is employed on 1 www7.informatik.tu-muenchen.de/ NEUROSEM/ different communicative levels. The variablility of aspects of the communicative situation a particle can refer to causes the apparent multifunctionality of the class. In particular, the content can refer to the following communicative domains: the speaker’s mental state, the hearer’s supposed mental state, the propositional level, the speaker-hearer interaction level, as well as the action level (Fischer, 1996b). It is assumed that the occurrence of each discourse particle has a specific reading in a natural dialogue. A number of such discourse functions has been identified in hypothesis-test cycles on four German corpora. The basic idea is that although many different features are involved on many different communicative levels, there is categorical perception involved in the interpretation of discourse particles. I.e. discourse functions are proposed to be more or less stable feature bundles of surface and pragmatic properties (Fischer, 1996a; Fischer, 1996b). The inventory of these discourse functions is supposed to be valid for all discourse particles, defining these linguistic elements from a functional perspective. As an example for different, yet related readings of a discourse particle consider the following two occurrences of English yes : Both are reactions to a proposal. In the first case, the semantic content of yes “you and I think the same” refers to the propositional level, signalling agreement on the subject in question as an answer particle. In the second example, yes is used to signal basic agreement to the communication partner ALTHOUGH the speaker has to reject the proposal. Here the semantic content refers to the interaction level. Example 1 yes, that would suit me. Example 2 yes, it is problematic because of the holidays. Themeanings of discourse particles are consequentlymodeledwith a fixed lexical core and additional context-dependent features which refer to different levels of communication. This renders the postulation of several polysemous items superfluous, and exploits context-dependent systematic ambiguity instead. Occurrences of discourse particles can be characterized further according to the specific functions they fulfillwith respect to the turn-taking system, the speech formulation and planning process, the discourse structure, as well as concerning their surface-level properties in utterances. For instance, yes can be used to take the turn as in example 3, to introduce a new topic (or even to open a conversation as in example 4) , and in utterance repairs (example 5). Example 3 yes, what would you suggest? Example 4 yes, hi hello my name is Quell. Example 5 oh yes, sorry, June. The individual features of description used concern both surface features (such as turn position: initial) and functions located in different pragmatic domains, for instance turn-taking or speech management functions. The corpus under consideration was recorded in a setting where one speaker had to teach another one to build a toyairplane (Sagerer et al., 1994). A simple example of a surface feature concerns the speaker’s role in uttering a discourse particle: for instance, it is more likely that a feedback signal is uttered by the constructor than the instructor. Other features, for example, concern the position of the particle in an utterance, as well as its combinations with other particles. Besides the functions a particle displays, an example of a pragmatic feature is constituted by the speech act the particle occurs in. It was found in distributional and functional analyses that there are significant correlations between particular dialogue acts, domain-specific speech acts of the preceding and the current utterance (Schmitz and Quantz, 1997), and the respective discourse particle (Fischer and Johanntokrax, 1995). Consequently, the preceding and current dialogue acts were coded for the feature-based description as pragmatic features. A catalogue of features and discourse functions is given in the appendix. To sum up, the linguistic features which are employed concern the many different functions of discourse particles in spoken language discourse, the different communicative domains the cognitive content of discourse particles refers to, as well as the surface realization of an occurrence. Principles of Linguistic Categorization Feature-based analysis has a long tradition in linguistics. Feature representations play a prominent role in most phonological theories, and they are also of considerable impact in theories of lexical meaning and grammatical categories. Looking at linguistic categorization from a cognitive perspective, we may emphasize the symbolic, i.e. binary (or n-ary) nature of representational features and the transient, context-dependent nature of category assignment. An exciting possibility from a neurocognitivepoint of view concerns the interpretation of linguistic units as truly perceptual categories on a par with visual image formation or auditory percepts (Scheler, 1997). Main analogies concern the existence of categorical perception, i.e. classification to stable higher-level units from lower-level descriptive input features, prototype formation, which concerns the abstraction from a large set of input patterns to a few central reference patterns, and contextual interpretation, which refers to the human brain’s ability to switch between various classifications of the same item depending on the perceived context. The system NEUROSEM has been developed with the goal of providing the specific tools necessary to perform connectionist semantic analysis for a wide range of applications. A precursor of NEUROSEMwas used formachine translationof aspectual categories and text correction of definite and indefinite articles (Scheler, 1996b; Scheler, 1996a). The main parts of NEUROSEM concern a flexible binary encoding scheme VGEN, an optional input tagger for surface categories (Bauer, 1995), a data analysis tool DATMAP, and a set of statistical and connectionist learning procedures. Clustering and Data Analysis Representation Issues and Sample Size According to the feature catalogue, a feature-based analysis was carried out on 150 randomly chosen occurrences of ja in a large German corpus (cf. (Sagerer et al., 1994)) and an additional set of 20 occurrences of oh. We arrive at a database of atomic features describing various aspects of the linguistic signal and its communicative setting. The goal of the computational model is (a) to provide an analysis of the data with respect to their stochastic and classificational properties and (b) to 0 20 40 60 80 100 120 0 10 20 30 40 50 60 70 80 90 100