Variational Graph Autoencoding as Cheap Supervision for AMR Coreference Resolution

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2790 - 2800 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Variational Graph Autoencoding as Cheap Supervision for AMR Coreference Resolution Irene Li1∗, Linfeng Song2†, Kun Xu2 and Dong Yu2 1Yale University, CT, USA irene.li@yale.edu 2Tencent AI Lab, Bellevue, WA, USA {lfsong,kxkunxu,dyu}@tencent.com Abstract Coreference resolution over semantic graphs like AMRs aims to group the graph nodes that represent the same entity. This is a cru- cial step for making document-level formal se- mantic representations. With annotated data on AMR coreference resolution, deep learn- ing approaches have recently shown great po- tential for this task, yet they are usually data hungry and annotating data is costly. We pro- pose a general pretraining method using vari- ational graph autoencoder (VGAE) for AMR coreference resolution, which can leverage any general AMR corpus and even automatically parsed AMR data. Experiments on bench- marks show that the pretraining approach achieves performance gains of up to 6% abso- lute F1 points. Moreover, our model signifi- cantly improves on the previous state-of-the- art model by up to 11% F1 points. 1 Introduction Abstract Meaning Representation (AMR) is a way to preserve the semantic meaning of a sentence in a graph (Banarescu et al., 2013). As shown in Figure 1, AMRs are directed and acyclic graphs where the nodes and edges indicate concepts and their seman- tic relations. As a sentence-level semantic repre- sentation, AMRs have been shown to be effective in many NLP tasks, including text summarization (Liu et al., 2015; Dohare et al., 2018), information extraction (Rao et al., 2017; Li et al., 2020b; Zhang and Ji, 2021), and machine translation (Song et al., 2019; Pham et al., 2020). More recently, the NLP tasks that are beyond the single-sentence level (Nallapati et al., 2016; Rajpurkar et al., 2016; Li et al., 2017; Chen et al., 2021) are attracting rising attention, and thus rep- resenting multiple sentences with AMR becomes important. To expand AMRs to represent multiple ∗Work done as an intern at Tencent AI Lab. † Corresponding author. and depart-01 shop receive-01 call-02 they op2 op1 ARG1 ARG0 ARG1 ARG0 ARG1 mod night last shop mod mod report-01 police ARG0 time S1: The police reported that a robbery happened in the shop last night. S2: They received the emergency call and departed immediately. emergency immediately manner Figure 1: An example of multi-sentence AMR coref- erence resolution. It contains two coreference clusters, marked by blue and pink respectively: police in S1 and They in S2; shop in S1 and the implicit mention of shop (with dashed edge and node) in S2. sentences, the task of AMR coreference resolution (O’Gorman et al., 2018) has been proposed, aiming at recognizing the concepts from multiple AMRs that represent the same entity. Figure 1 illustrates the AMR graphs of two consecutive sentences in a news article. Given them as the input, an AMR coreference resolver needs to group police and they (colored with blue), …

Content truncated to 500 words