This paper presents a predicate-argument structure (PAS) analysis for dialogue systems in Japanese. Conventional PAS analyses have been applied to newspaper articles; however, there are differences between newspapers and dialogues. Therefore, to comprehensively deal with these differences, we constructed a PAS analyzer for dialogues as a type of domain adaptation from newspapers. Because pronominalization and ellipses frequently appear in dialogues, we utilized a strategy that simultaneously resolves zero-anaphora and adapts our PAS analyzer to dialogues. By incorporating parameter adaptation and automatically acquiring knowledge from large text corpora, we performed a PAS analysis that is specific to dialogues. Our PAS analyzer has a higher accuracy compared with the analyzer for newspaper articles.

View full abstract

We propose a novel task that identifies cross-document sentence relations from document pairs. Although there are numerous studies that focus on finding sentence relations from just one document or conversation, only few studies are proposed for cross-documents. Examples of cross-document sentence relations are question–answer relations, request–response relations, and so on. Finding such relations will lead to many applications since the cross-document sentence relations are useful to explain document-based conversations on a more fine-grained level. For instance, we can extract communications from cross-documents by accumulating sentences having relations. To detect such relations, we regard this task as the classification problem and employ the conditional random fields. In particular, we modify a previous method that focuses on finding relations from conversations using sentence types to our task. Furthermore, we propose a combined model that simultaneously estimates sentence types and relations. The experiments are performed on review and reply on an internet service for hotel reservation, and the results show that our proposed model achieves 46.6% precision and 61.0% recall, which outperforms previous models.

View full abstract