A Model for Analyzing Structures of Coherence Relations Using Features of Verbs and Subjects

Abstract

The model proposed in this paper analyzes meanings of coherence relations using features of verbs and subjects in Japanese complex sentences, and then analyzes structures of coherence relations using the meanings of coherence relation. Dependency structures of Japanese complex sentences are usually analyzed using hierarchical classification of conjunctions and conjunctive particles. But, conjunctions and conjunctive particles usually have multiple senses and are ambiguous. If a conjunction or conjunctive particle in a subordinate clause has a different sense, the subordinate clause may modify a predicate in a different clause. So, the model analyzes the coherence relations between subordinate clauses and a main clause using the features of verbs and subjects, and defines the meanings of coherence relations. Then the meanings of coherence relations are classified according to distance of coherence. The model uses this classification of coherence relations to analyze the structures of coherence relations. Volition, aspect, mood, voice and semantic category etc.are used as the features of verbs, and animate or inanimate etc. is used as the features of subjects. The model is evaluated by examples from actual documents, and shows 98.4% accuracy. Since the model using the classification of conjunctions and conjunctive particles shows 97.0% accuracy with the same examples, the model proposed in this paper, decreases the error rate by half.