DRTS Parsing with Structure-Aware Encoding and Decoding

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6818–6828 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 6818 DRTS Parsing with Structure-Aware Encoding and Decoding Qiankun Fu1,2,3, Yue Zhang2,3, Jiangming Liu4 and Meishan Zhang5 1. Zhejiang University 2. School of Engineering, Westlake University 3. Institute of Advanced Technology, Westlake Institute for Advanced Study 4. ILCC, University of Edinburgh 5. School of New Media and Communication, Tianjin University, China fqiankun@gmail.com and zhangyue@westlake.edu.cn jiangming.liu@ed.ac.uk and mason.zms@gmail.com Abstract Discourse representation tree structure (DRTS) parsing is a novel semantic parsing task which has been concerned most recently. State-of-the-art performance can be achieved by a neural sequence-to-sequence model, treating the tree construction as an incremen- tal sequence generation problem. Structural information such as input syntax and the intermediate skeleton of the partial output has been ignored in the model, which could be potentially useful for the DRTS parsing. In this work, we propose a structural-aware model at both the encoder and decoder phase to integrate the structural information, where graph attention network (GAT) is exploited for effectively modeling. Experimental results on a benchmark dataset show that our proposed model is effective and can obtain the best performance in the literature. 1 Introduction Discourse representation tree structure (DRTS) is a form of discourse structure based on Discourse Representation Theory of Kamp and Reyle (1993), a popular theory of meaning representation (Kamp, 1981; Asher, 1993; Asher and Lascarides, 2003). It is designed to account for a variety of linguistic phe- nomena, including the interpretation of pronouns and temporal expressions within and across sen- tences. Correspondingly, as one type of discourse parsing, DRTS parsing (Liu et al., 2018) can be helpful for paragraph or document-level text un- derstanding by converting DRS to tree-style DRTS. (Liu et al., 2019). Figure 1 shows an example of DRTS, where the leaf nodes are discourse representation units (DRUs), upon which a discourse tree structure built. In particular, a DRU consists of several individ- ual tuples, where each tuple denotes a relation in- side the DRU. For example, there is a relationship Figure 1: Left: Our proposed model with two structure- aware module. Right: The DRTS for a clause in a document: “The letterx4 warns Jewish womenx16 that they will suffer if they date Arab men. p4” “That” between the specific entity x16 and a propo- sition p4. The relationships between the DRUs are organized by a tree skeleton, which includes three types of nodes: the S(DRS) nodes to introduce DRU, the relation nodes for inter-DRU relation- ship, and the variable nodes, which are used to define S(DRS) (e.g., p4, k1 and k4 ). There have been only a few existing stud- ies related to DRTS parsing (van Noord et al., 2018a,b). In particular, the end-to-end encoder- decoder model of Liu et al. (2019) gives the state- of-the-art performance, which converts the task into a sequence-to-sequence problem. The input sequence consists of words in paragraphs, encoded by a BiLSTM structure, and the output sequence is top-to-bottom …

Content truncated to 500 words