Acquisition and Analysis of Distributed Representations of Words and Sentences Focusing on Language Learners' Errors

2023 Fiscal Year Final Research Report

• Project/Area Number: 19KK0286
• Research Category: Fund for the Promotion of Joint International Research (Fostering Joint International Research (A))
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 61030:Intelligent informatics-related
• Research Institution: Hitotsubashi University (2023); Tokyo Metropolitan University (2019-2022)
• Principal Investigator: KOMACHI Mamoru 一橋大学, 大学院ソーシャル・データサイエンス研究科, 教授 (60581329)
• Project Period (FY): 2020 – 2023
• Keywords: 自然言語処理 / 深層学習 / 文法誤り訂正 / 言語学習支援 / 言語教育支援

Outline of Final Research Achievements

We have worked on the construction of datasets for grammatical error correction in English, Japanese, and Chinese, as well as the analysis and comprehensive evaluation of outputs from multilingual grammatical error correction systems using deep learning. Below is an overview of the research achievements conducted throughout the research period: (1) Application of pre-trained models to grammatical error correction, (2) Acceleration of grammatical error correction systems, (3) Proposal of a grammatical error correction method using a pseudo-learner corpus considering learners' errors, (4) Analysis and improvement of the diversity of grammatical error correction outputs, (5) Transfer learning of language knowledge for grammatical error correction using multilingual models, and (6) Development and dataset construction of an automatic evaluation method for grammatical error correction

Free Research Field

自然言語処理

Academic Significance and Societal Importance of the Research Achievements

本研究を通じて、最先端の深層学習を用いた文法誤り訂正手法の到達点と限界について明らかになりました。文法誤り訂正の性能が見違えるように改善された一方、これまで用いられてきた文法誤り訂正の評価データセットが深層学習時代の文法誤り訂正手法の評価には適さないことが明らかになり、多言語での評価用のデータセットの構築や、それらを用いた適切な評価尺度の開発の必要性が示唆され、言語学習者の誤用の評価の重要性が再確認されました。