DCZAR: ゼロ照応解析に基づく項省略補完による対話応答生成

<p>Human conversation attempts to build common ground consisting of shared beliefs, knowledge, and perceptions that form the premise for understanding utterances. Recent deep learning–based dialogue systems use human dialogue data to train a mapping from a dialogue history to responses, but common ground not directly expressed in words makes it difficult to generate coherent responses by learning statistical patterns alone. Inspired by the idea of zero anaphora resolution (ZAR), we propose Dialogue Completion using Zero Anaphora Resolution (DCZAR), a framework that explicitly completes omitted information in a dialogue history and generates responses from the completed history. The DCZAR framework consists of three models: a predicate-argument structure analysis (PAS) model, a dialogue completion (DC) model, and a response generation (RG) model. The PAS model analyzes the omitted arguments (zero pronouns) in the dialogue, and the DC model determines which arguments to complete and where to complete them and explicitly completes the omissions in the dialogue history. The RG model, trained by the complementary dialogue history and response pairs, generates a response. The PAS and RG models are constructed by fine-tuning the common pretrained model with a dataset corresponding to each task, while the DC model uses a pretrained model without fine-tuning. We used the Japanese Wikipedia dataset and Japanese postings to Twitter to build our pretrained models. Since tweets are like dialogues in that they contain many abbreviations and short sentences, the model pretrained with tweets is expected to improve the performance of ZAR and dialogue response generation. Experimental results show that the DCZAR framework can be used to generate more coherent and engaging responses. Analysis of the responses shows that the model generated responses that were highly relevant to the dialogue history in dialogues with many characters.</p>