Text semantic parsing combining dynamic knowledge obtained from preceding context and static knowledge obtained in pre-training

2022 Fiscal Year Final Research Report

• Project/Area Number: 19K12112
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 61030:Intelligent informatics-related
• Research Institution: Tohoku University
• Principal Investigator: Matsubayashi Yuichiroh 東北大学, 教育学研究科, 准教授 (20582901)
• Project Period (FY): 2019-04-01 – 2023-03-31
• Keywords: 省略解析 / 述語項構造 / 意味解析 / 文章理解

Outline of Final Research Achievements

The performance of omission analysis, which is fundamental to "accurate reading" of sentences, had a large gap compared to general human reading comprehension ability. The goal of this research was to improve the conventional semantic parsing technique, which only looks at a few sentences around the target sentence, by (1) constructing a parsing model that understands meaning based on accumulating the meaning of sentences in the preceding context (= dynamic knowledge), (2) establishing an effective and efficient method to express the common sense knowledge required for inference (= static knowledge), and (3) realizing the natural inference based on the combination of such dynamic and static knowledge. As a result, we were successful in implementing an omission analysis system that significantly outperformed previous methods' performance. Our proposed system, developed in this project, has been publicly released as open source software.

Free Research Field

自然言語処理

Academic Significance and Societal Importance of the Research Achievements

省略解析は文章の意味を正確に理解するAIの実現に不可欠な要素であり、日本語解析のボトルネックとなっていたこの基盤技術の解析精度向上により、応用技術の発展可能性が増大した意義は大きい。開発したシステムは一般公開し、実世界テキスト解析に適用可能である。精度向上の鍵となったアイデアは、汎用的言語モデルに対して学習の形態を大きく変更することなくシームレスに意味解析の能力を増強するものであり、その他の言語処理技術の性能向上に対しても応用可能性を秘めている。加えて、研究過程で得られた知見から書き手の省略判断分析という新たな研究の方向性を展開し、教育応用等へのシードを生んだ点も学術的意義として挙げられる。