Dependency Patterns of Complex Sentences and Semantic Disambiguation for Abstract Meaning Representation Parsing

Proceedings of the 10th Conference on Lexical and Computational Semantics, pages 212–221 August 5–6, 2021, Bangkok, Thailand (online) ©2021 Association for Computational Linguistics 212 Dependency Patterns of Complex Sentences and Semantic Disambiguation for Abstract Meaning Representation Parsing Yuki Yamamoto* Yuji Matsumoto** Taro Watanabe* *Nara Institute of Science and Technology {yamamoto.yuki.yt0, taro}@is.naist.jp **RIKEN Center for Advanced Intelligence Project {yuji.matsumoto}@riken.jp Abstract Abstract Meaning Representation (AMR) is a sentence-level meaning representation based on predicate argument structure. One of the challenges we find in AMR parsing is to capture the structure of complex sentences which ex- presses the relation between predicates. Know- ing the core part of the sentence structure in ad- vance may be beneficial in such a task. In this paper, we present a list of dependency patterns for English complex sentence constructions de- signed for AMR parsing. With a dedicated pat- tern matcher, all occurrences of complex sen- tence constructions are retrieved from an input sentence. While some of the subordinators have semantic ambiguities, we deal with this problem through training classification mod- els on data derived from AMR and Wikipedia corpus, establishing a new baseline for future works. The developed complex sentence pat- terns and the corresponding AMR descriptions will be made public1. 1 Introduction Abstract Meaning Representation (AMR) is a sentence-level meaning representation based on predicate argument structure (Banarescu et al., 2013). AMR Parsing is the task of transforming a sentence into an AMR graph with nodes and edges, each representing a concept or relation. While early studies (Flanigan et al., 2014; Wang et al., 2015; Artzi et al., 2015; Pust et al., 2015) used de- pendency parsers to integrate syntactic features to their models, recent deep neural network-based ap- proaches (Konstas et al., 2017; Peng et al., 2017; Zhang et al., 2019; Cai and Lam, 2020) tend to en- code the input sentence as a sequence without con- sidering its syntactic structure. Generally speaking, syntactic and semantic structures share much in common. It is assumed 1Code and resource are available at https://github. com/yama-yuki/skeletal-amr. �� ���� �� �� �� ����� ����� ������ ����� �� ���� �� ������ �� ���������� �� �� ���� ���������� ���� �� ���� Figure 1: Representation of as long as-construction in dependency tree (left) and AMR graph (right). that dependency trees and semantic role label- ing structures have a strong correlation in that nsubj and dobj can be used interchangeably for ARG0 and ARG1 role (Xia et al., 2019). Since AMR is annotated based on PropBank frames (Palmer et al., 2005), the same could be said for AMR structures. This holds to be true for a simple sentence, which is basically a matrix clause, comprised of a predi- cate and its arguments. However, it is not always the case with complex sentence constructions, each of which consists of a matrix clause and one or more subordinate clause(s). Consider Figure 1 which shows both dependency and AMR represen- tation of a complex sentence with a subordinator as long as. While variables S’s and V’s are inter- changeable …

Content truncated to 500 words