Learning conversation agents for long-vision response generation in multi-round communications

2021 Fiscal Year Final Research Report

• Project/Area Number: 19K20345
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 61030:Intelligent informatics-related
• Research Institution: The University of Tokushima
• Principal Investigator: KANG Xin 徳島大学, 大学院社会産業理工学研究部(理工学域), 助教 (80777350)
• Project Period (FY): 2019-04-01 – 2022-03-31
• Keywords: 対話生成 / 対話完成程度 / 対話完成効率 / 対話ユーザー満足度

Outline of Final Research Achievements

Dialogue system has been an important research field in intelligent Human-Computer Interface (HCI) studies and has been widely applied as an intelligent service in smart products and self-driving systems. Traditional dialogue system consists of four independent modules, which are the utterance understanding module, the dialogue-state updating module, the dialogue policy managing module, and the utterance generating module. Such system is difficult to tune because independent modules cannot broadcast the learning error between each other and turns to accumulate prediction errors through the pipeline. In this project, the applicant proposed a topic knowledge driven end-to-end neural dialogue system. Specifically, the applicant proposed an automatic dialogue topic extraction method, an automatic dialogue quality evaluation method, and a generative adversarial network for high quality dialogue generation, and a knowledge-based nurse-patient dialogue system.

Free Research Field

自然言語処理

Academic Significance and Societal Importance of the Research Achievements

モジュール連結に基づく対話システムにより，本研究はモジュール機能をエンドツーエンド深層学習に統合して，話題知識抽出，対話分品質評価と敵対的生成学習を用いて，マルチターン対話のタスク完成程度，完成効率とユーザー満足度を向上し，対話を用いた知能ヒューマンマシンインターフェイス技術を構築する上に，スマート製品におけるユーザー体験や学習コストの改善や医療介護支援ロボット開発の推進を期待されている．