This paper develops a general approach to contextual reasoning in natural language processing. Drawing on the view of natural language interpretation as abduction (Hobbs et al., 1993), we propose that interpretation provides an explanation of how an utterance creates a new discourse context in which its interpreted content is both true and prominent. Our framework uses dynamic theories of semantics and pragmatics, formal theories of context, and models of attentional state. We describe and illustrate a Prolog implementation. 1 Problem Statement Context in discourse and dialogue involves at least two components, shared information and coordinated attention; each of these components suggests a mechanism through which speakers may frame natural language utterances to fit the context. Shared information makes available to the par-ticipants a body of facts that characterize the world under discussion (here we mean to be neutral between various characterizations of this availability in terms of common ground, common knowledge, mutual belief, etc.); accordingly, language users may presume that information in an utterance links up with this body of facts, whenever possible (Lascarides et al., 1992; Hobbs et al., 1993). Coordinated attention puts certain entities at the forefront of the discussion; accordingly, language users may presume that descriptions in an utterance refer to these entities, whenever possible (Grosz and Sidner, 1986; Grosz et al., 1995; Walker et al., 1997). How these two potential mechanisms interact is an empirical question—but this question can only be addressed within a framework that describes in

Contents 1 Introduction 1 2 Model Building 3 2.1 First-Order Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 2.2 Constructing Models for Logical Theories . . . . . . . . . . . . . . . . . . . . . . . 4 2.3 Inconsistent Theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 3 Linguistic Applications 5 3.1 Information Seeking Dialogue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 3.2 Controlling a Mobile Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 3.3 Question Answering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 4 Using Inference Tools 8 4.1 Automated Model Builders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 4.2 Automated Theorem Provers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 4.3 System Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10<F18.67

Model generation calculi, close relatives of tableau calculi for theorem proving, can be used as competence models for semantic natural language understanding. Unfortunately, existing model generation calculi are not yet plausible as performance models of actual human processing, since they fail to capture computational aspects of human language processing.

In recent work, tableau-based model generation calculi have been used as computational models of the reasoning processes involved in utterance interpretation. In this linguistic application of an inference technique that was originally developed for automated theorem proving, natural language understanding is treated as a process of generating Herbrand models for the logical form of an utterance in a discourse. This approach captures anbiguity by generating multiple models for input logical forms.

ABSTRACT: In this article we discuss what constitutes a good choice of semantic representation, compare different approaches to constructing semantic representations for fragments of natural language, and give an overview of recent menthods for employing inference engines for natural understanding tasks.

Automatic knowledge acquisition is one of the bottlenecks in artificial intelligence and large-scale applications of natural language processing (NLP). There are many efforts to create large knowledge bases (KBs) or to automatically derive knowledge from large text corpora. On the one hand, we meet KBs like CYC, where a tremendous amount of work has been invested by knowledge enterers who have manually formalized large stocks of knowledge. The other extreme are projects using flat (mostly statistically based) methods for extracting knowledge from texts. These techniques seldom produce results with a clear semantic interpretation and sufficient quality for NLP applications, however. MAC-QUIK is a project to automatically acquire knowledge from natural language sources (like text corpora or lexicons) by means of a deep syntacticosemantic analysis and subsequent assimilation of the generated representations into a coherent KB. The paper emphasizes the role of a homogeneous formalism for interfacing between NLP and inferential question answering, and it demonstrates its use for a deductive treatment of coreference resolution. 1

We present an architecture for spoken dialogue systems where first-order inference (both theorem proving and model building) plays a crucial role in interpreting utterances of dialogue participants and deciding how the system should respond and carry out instructions. The dialogue itself is represented as a DRS which is translated into first-order logic for inference tasks. The system is implemented as a society of OAA-agents, and evaluated against a specific application (home automation).

Bridging reference resolution presupposes world knowledge coded in some declarative form as well as inferencing methods capable of reasoning on the basis of this knowledge. Furthermore, in order to make bridging reference resolution to become feasible and practicable, the availability of domain knowledge at a large scale as well as of powerful and robust inferencing techniques seems crucial. In this paper I attempt to show how bridging reference resolution can be made feasible by exploiting ontologies developed within the context of the Semantic Web as well as model building techniques. For this purpose I present a DRT-based approach combining these two very promising elements. 1 Introduction In this paper I understand bridging in line with Asher et al. ([2]) as "the inference that two objects or events that are introduced in a text are related in a particular way that isn't explicitly stated". Here follow some examples 1 taken from Clark ([9]): Example 1.1 I walked into the room. The chandeliers sparkled brightly. Example 1.2 I met two people yesterday. The woman told me a story. Example 1.3 John was murdered yesterday. The murderer got away.

Abstract not found