K-CSRL: Knowledge Enhanced Conversational Semantic Role Labeling

Abstract

Semantic role labeling (SRL) is widely used to extract predicate-argument pairs from sentences. Traditional SRL methods can perform well on the single sentence but fail to work in dialogue scenario where ellipsis and anaphora frequently occurs. Some research work has been proposed to solve this problem, i.e. Conversational Semantic Role Labeling (CSRL), but there are still huge room for improvements. The error case study of BERT-based CSRL model has shown that the majority of the errors are observed in boundary matching, especially in entity mention detection. We think the premier cause of this kind of error is the deficiency of external knowledge such that the ill-informed model cannot correctly capture and correlate the entities. To this end, we propose to incorporate external knowledge into BERT using visible masking strategy. We evaluate our proposed model on DuConv dataset. Experimental results show that our model with knowledge enhancement outperforms the benchmarks. Further analysis also demonstrates that dialogue SRL can benefit from external knowledge.