Testing Gricean Constraints on a WordNet-based Coherence Evaluation System

This paper presents a computational method for analyzing Gricean constraints for the purpose of evaluating text coherence. Our system consists of a knowledge base constructed on top of WordNet, an inference engine that establishes discourse semantic paths between the input concepts and a mechanism of relating them to the context. Grice's maxims provide conditions to test coherence, while the semantic paths provide the space on which these conditions are tested. The computational method is based on a marker-propagation technique that is independent of the size of the knowledge base. The paper describes the method and provides results obtained with the system. The key idea of this paper is to link Grice's maxims for conversational logic with high level inferences drawn from a linguistic knowledge base. We discuss how the Gricean maxims may be tested in a real natural language processing system and how the information extracted plays a role in the analysis of context, dialogue coherence, and intentions. A text is considered to be coherent when it is logically connected and intelligible. The logical connections are semantic relationships between concepts. We are experimenting with a text understanding system that operates on a large linguistic knowledge base with concepts and relations between concepts. Inference rules are constructed by chaining semantic relations in the knowledge base. A few of these rules lead to deductions , but many provide only plausible inferences. Semantic paths between concepts may be established as patterns of inference rules. The information along such paths provides an explanation for the relationship between two concepts. In this paper, we link these semantic paths to Gricean constraints. The concepts and the relations that form the semantic paths provide context and information about text coherence. The Gricean constraints are rst tested by analyzing the paths; and this provides information to validate or lter out irrelevant inferences. The path nding algorithm designed by us uses a marker propagation paradigm that is highly parallel and asynchronous. Unlike other attempts that use marker-passing (Charniak 1986), (Norvig 1987), we use propagation rules that guide the movement of markers along some selected relations. Any marker present on a node may further propagate along semantic relations that (a) generalize, (b) deene the features or (c) entail or are entailed by the properties of the concept represented by that node. Whenever two markers originated from diierent nodes of the knowledge base reach the same node, a collision of the …