Semantic Frame Induction with Deep Metric Learning

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1833–1845 May 2-6, 2023 ©2023 Association for Computational Linguistics Semantic Frame Induction with Deep Metric Learning Kosuke Yamada1 Ryohei Sasano1,2 Koichi Takeda1 1Graduate School of Informatics, Nagoya University, Japan 2RIKEN Center for Advanced Intelligence Project, Japan yamada.kosuke.v1@s.mail.nagoya-u.ac.jp, {sasano,takedasu}@i.nagoya-u.ac.jp Abstract Recent studies have demonstrated the useful- ness of contextualized word embeddings in un- supervised semantic frame induction. However, they have also revealed that generic contextu- alized embeddings are not always consistent with human intuitions about semantic frames, which causes unsatisfactory performance for frame induction based on contextualized em- beddings. In this paper, we address supervised semantic frame induction, which assumes the existence of frame-annotated data for a subset of predicates in a corpus and aims to build a frame induction model that leverages the an- notated data. We propose a model that uses deep metric learning to fine-tune a contextual- ized embedding model, and we apply the fine- tuned contextualized embeddings to perform semantic frame induction. Our experiments on FrameNet show that fine-tuning with deep metric learning considerably improves the clus- tering evaluation scores, namely, the B-CUBED F-SCORE and PURITY F-SCORE, by about 8 points or more. We also demonstrate that our approach is effective even when the number of training instances is small. 1 Introduction Semantic frames are knowledge resources that re- flect human intuitions about various concepts such as situations and events. One of the most represen- tative semantic frame resources is FrameNet (Baker et al., 1998; Ruppenhofer et al., 2016), which consists of semantic frames, lexical units (LUs) that evoke these frames, and collections of frame- annotated sentences. Semantic frame induction is the task of grouping predicates, typically verbs, ac- cording to the semantic frames they evoke.1 For example, given the verbs in the example sentences 1Strictly speaking, this task can be divided into two sub- tasks: verb clustering, which groups verbs according to the frames that they evoke, and argument clustering, which groups arguments of verbs according to their roles. In this study, we focus the former, the verb clustering task. Frame Example sentence FILLING (1) She covered her mouth with her hand. (2) I filled a notebook with my name. PLACING (3) You can embed graphs in your worksheet. (4) He parked the car at the hotel. REMOVING (5) Volunteers removed grass from the marsh. (6) They’d drained the drop from the teapot. TOPIC (7) Each database will cover a specific topic. (8) Chapter 8 treats the educational advantages. Table 1: Example sentences and the frames that their verbs evoke in FrameNet. (1) (a) Vanilla BERT (b) Fine-tuned BERT w/ AdaCos (2) (3) (4) (6) (7) (8) (5) Figure 1: 2D t-SNE projections of the contextualized embeddings of verbs annotated with either the FILL- ING (•), PLACING (×), REMOVING (■), or TOPIC (+) frames in FrameNet, by using (a) vanilla BERT or (b) fine-tuned BERT with AdaCos. The numbers in paren- theses correspond to the examples listed in Table 1. …

Content truncated to 500 words