End-to-End AMR Coreference Resolution

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4204–4214 August 1–6, 2021. ©2021 Association for Computational Linguistics 4204 End-to-End AMR Coreference Resolution Qiankun Fu1,2,3, Linfeng Song4, Wenyu Du2,3, Yue Zhang2,3 1. Zhejiang University 2. School of Engineering, Westlake University 3. Institute of Advanced Technology, Westlake Institute for Advanced Study 4. Tencent AI Lab, Bellevue, WA, USA fqiankun@gmail.com and lfsong@tencent.com {duwenyu, zhangyue}@westlake.edu.cn Abstract Although parsing to Abstract Meaning Rep- resentation (AMR) has become very popu- lar and AMR has been shown effective on many sentence-level tasks, little work has stud- ied how to generate AMRs that can repre- sent multi-sentence information. We introduce the first end-to-end AMR coreference resolu- tion model in order to build multi-sentence AMRs. Compared with the previous pipeline and rule-based approaches, our model allevi- ates error propagation and it is more robust for both in-domain and out-domain situations. Be- sides, the document-level AMRs obtained by our model can significantly improve over the AMRs generated by a rule-based method (Liu et al., 2015) on text summarization. 1 Introduction Abstract Meaning Representation (AMR) (Ba- narescu et al., 2013) is a semantic formalism for natural language understanding. It represents a sentence as a rooted, directed and acyclic graph, where nodes (e.g., “Bill” in Figure 1) represents concepts and edges (e.g., “:arg0”) are the seman- tic relations. Encompassing knowledge of named entities, semantic roles and coreference structures, AMR has been proven effective for downstream tasks, including information extraction (Rao et al., 2017), text summarization (Liu et al., 2015; Hardy and Vlachos, 2018; Liao et al., 2018), paraphrase detection (Issa Alaa Aldine et al., 2018), event de- tection (Li et al., 2015), machine translation (Song et al., 2019b) and dialogue understanding (Bonial et al., 2020). Existing work on AMR mainly focuses on in- dividual sentences (Lyu and Titov, 2018; Naseem et al., 2019; Ge et al., 2019; Zhang et al., 2019; Cai and Lam, 2020a; Zhou et al., 2020). On the other hand, with the advance of neural networks in NLP, tasks involving multiple sentences with leave-11 person name name city Bill Paris Sentence1: Bill left for Paris. Sentence2: He arrived at noon. arrive-01 Paris he date-entity noon :name :name :arg0 :arg3 :arg1 :arg2 :dayperiod :op1 :op1 Figure 1: Multi-sentence AMR example, where nodes with the same non-black color are coreferential and the dotted ellipse represents an implicit role coreference. cross-sentence reasoning (e.g., text summarization, reading comprehension and dialogue response gen- eration) have received increasing research atten- tion. Given the effectiveness of AMR on sentence- level tasks (Pan et al., 2015; Rao et al., 2017; Issa Alaa Aldine et al., 2018; Song et al., 2019b), it is important to extend sentence-level AMRs into the multi-sentence level. To this end, a prerequisite step is AMR coreference resolution, which aims to find the AMR components referring to the same entity. Figure 1 shows the AMR graphs of two consecutive sentences in a document. An AMR coreference resolution model need …

Content truncated to 500 words