EFSP-TE: End-to-End Frame-Semantic Parsing with Table Encoder

Abstract

Frame-Semantic Parsing (FSP) aims to extract frame-semantic structures from text. The task usually involves three subtasks sequentially: Target Identification (TI), Frame Identification (Fl), and Frame Semantic Role Labeling (FSRL). The three subtasks are closely related while most previous studies model them individually, encountering error propagation and running efficiency problems. Recently, an end-to-end graph-based model is proposed to jointly process three subtasks in one model. However, it still encounters three problems: insufficient semantic modeling between targets and arguments, span missing, and lacking knowledge incorporation of FrameNet. To address the mentioned problems, this paper presents an End-to-end FSP model with Table Encoder (EFSP-TE), which models FSP as two semantically dependent region classification problems and extracts frame-semantic structures from sentences in a one-step manner. Specifically, EFSP-TE incorporates lexical unit knowledge into context encoder via saliency embedding, and develops an effective table representation learning method based on Biaffine network and multi-layer ResNet-style-CNNs (Convolutional Neural Networks), which can fully exploit word-to-word interactions and capture the information of various levels of semantic relations between targets and arguments. In addition, it adopts two separate region-based modules to obtain potential targets and arguments, followed by two interactive classification modules to predict the frames and roles for the potential targets and arguments. Experiments on two public benchmarks show that the proposed approach achieves state-of-the-art performance in end-to-end setting.