Span-Based Semantic Role Labeling with Contrastive Learning
Abstract: Contrastive learning is widely recognized for its ability to understand the relationships between data and map them into a high-dimensional feature space. In this study, we apply this technique to semantic role labeling, constructing a model that effectively captures the relationships between spans and labels and determines spans accurately. Our model integrates the characteristics of both a conventional span-based model, which predicts spans for labels, and a model that is comparable to state-of-the-art, which predicts labels for spans. In our experiments, we apply these models to NPCMJ-PT, a Japanese corpus that is annotated with semantic role labels and has about 52,500 entries. The semantic roles are defined with 32 types of labels such as Arg0, Arg1 and ArgM-LOC, which are similar to PropBank. The experimental results show that our model outperforms the conventional span-based models, achieving a highest F1 score of 81.2.
Paper Type: Long
Research Area: Syntax: Tagging, Chunking and Parsing
Research Area Keywords: semantic parsing
Contribution Types: NLP engineering experiment
Languages Studied: Japanese
Submission Number: 2506
Loading