Encoder-Decoder Based Unified Semantic Role Labeling with Label-Aware Syntax

Currently the unified semantic role labeling (SRL) that achieves predicate identification and argument in an end-to-end manner has received growing interests. Recent works show leveraging syntax knowledge significantly enhances SRL performances. In this paper, we investigate a novel framework based on sequence-to-sequence architecture with double enhancement both encoder decoder sides. side, propose label-aware graph convolutional network (LA-GCN) to encode syntactic dependent arcs labels into BERT-based word representations. creatively design pointer-network-based model for detecting predicates, arguments roles jointly. Our pointer-net is able make decisions by consulting all input elements global view, meanwhile it syntactic-aware incorporating information from LA-GCN. Besides, high-order interacted attention introduced previously recognized triplets help current decision. Empirical experiments our outperforms existing graph-based methods CoNLL09 Universal Proposition Bank datasets. In-depth analysis demonstrates can effectively capture correlations between structures.