2018 Fiscal Year Final Research Report

The computational modeling of deep, robust discourse analysis by integrating abductive reasoning, machine learning, and physical computing

Research Project

PDF

Project/Area Number	15H01702
Research Category	Grant-in-Aid for Scientific Research (A)
Allocation Type	Single-year Grants
Section	一般
Research Field	Intelligent informatics
Research Institution	Tohoku University
Principal Investigator	Inui Kentaro 東北大学, 情報科学研究科, 教授 (60272689)
Co-Investigator(Kenkyū-buntansha)	岡崎直観東京工業大学, 情報理工学院, 教授 (50601118) 井之上直也東北大学, 情報科学研究科, 助教 (80778605) 鈴木潤東北大学, 情報科学研究科, 准教授 (80396150)
Project Period (FY)	2015-04-01 – 2019-03-31
Keywords	自然言語処理 / 人工知能 / 知能情報学 / 談話解析 / 仮説推論 / 知識獲得 / 物理計算
Outline of Final Research Achievements	This project aimed at the computational modeling of deep, robust discourse analysis that was able to "read between the lines" by integrating abductive reasoning, machine learning, and physical computing. The main achievements are as follows. First, we significantly enhanced the computational capacity of abductive reasoning by formalizing the problem as weighted maximum satisfiability and devising efficient pruning methods. Second, we developed novel methods for large-scale knowledge acquisition from both Web and Wikipedia documents and demonstrated the impact of leveraging acquired knowledge on semantic and discourse analysis. Third, we built and empirically evaluated a computational model that innovatively integrated abductive reasoning and physical simulation to predict risks involved in given traffic scenes. The resources and tools developed in this project are made publicly available on our website.
Free Research Field	自然言語処理
Academic Significance and Societal Importance of the Research Achievements	本研究は、文章の「行間」を読むことができる計算モデルの構築を目指して論理推論と機械学習の融合を図ったもので、極めて独自性の高い試みであった。仮説論理推論の手法についても既存手法に比べて百倍以上の速度改善を達成しており、実世界データにおける仮説推論の実行可能性を飛躍的に引き上げた。談話解析を物理計算にグラウンディングさせる研究も先例のない挑戦的な試みであったが、その成果として運転シーンからの危険予測技術など実世界へ直接的に組み込める技術の雛形を開発し、「仮説推論による意図理解」というまったく新しい方向性の技術の実例を示した。