Building Semantic Frames using Contextualized Word Embeddings

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K12012
• 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: Nagoya University
• Principal Investigator: Sasano Ryohei 名古屋大学, 情報学研究科, 准教授 (70603918)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: 意味フレーム / 文脈化単語埋め込み

Outline of Final Research Achievements

In this research project, we worked mainly on automatically constructing semantic frame knowledge using contextualized word embeddings. Specifically, we proposed a method that uses masked word embedding and two-stage clustering for the frame induction task, in which verbs are clustered according to the frames they evoke, and a method that performs fine-tuning of the contextualized word embedding model based on deep metric learning followed by clustering of the verbs. The proposed method achieves a higher performance than the existing methods. In frame element knowledge acquisition task, we proposed a method that performs fine-tuning of the contextualized word embedding model based on deep metric learning followed by clustering of agruments of frame evoking words, and achieved a higher performance than the existing methods.

Free Research Field

自然言語処理

Academic Significance and Societal Importance of the Research Achievements

本研究の成果は、これまで人手で行われてきた意味フレーム資源の開発コストの低減する可能性がある。具体的には、これまで大規模な意味フレーム資源が整備されてこなかった言語を対象に意味フレームを自動構築したり、人手によるフレーム資源の整備の補助に用いることが可能である。また、大規模なコーパスから学習された大規模言語モデルが、人が言語理解の前提として持つ経験的知識をどの程度含んでいるかの分析に貢献する可能性がある。