UofR at SemEval-2016 Task 8: Learning Synchronous Hyperedge Replacement Grammar for AMR Parsing

Proceedings of SemEval-2016, pages 1185–1189, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics UofR at SemEval-2016 Task 8: Learning Synchronous Hyperedge Replacement Grammar for AMR Parsing Xiaochang Peng and Daniel Gildea Department of Computer Science University of Rochester Rochester, NY 14627 Abstract In this paper, we apply a synchronous-graph- grammar-based approach to SemEval-2016 Task 8, Meaning Representation Parsing. In particular, we learn Synchronous Hyperedge Replacement Grammar (SHRG) rules from aligned pairs of sentences and AMR graphs. Then we use Earley algorithm with cube- pruning for AMR parsing given new sentences and the learned SHRG. Experiments on the evaluation dataset demonstrate that competi- tive results can be achieved using a SHRG- based approach. 1 Introduction Abstract Meaning Representation (AMR) (Ba- narescu et al., 2013) is a semantic formalism where the meaning of a sentence is encoded as a rooted, directed graph. Figure 1 shows an example of the edge-labeled representation of an AMR graph, where the edges are labeled while the nodes are not. AMR utilizes PropBank frames, non-core seman- tic roles, coreference, named entity annotations and other semantic phenomena to represent the semantic structure of a sentence and abstracts away its syntax form. These properties render AMR representation useful in applications like question answering and semantics-based machine translation. SemEval-2016 Task 8 is the task of recovering this type of semantic formalism for plain text. A large corpus of annotated English/AMR pairs is pro- vided to learn this mapping. Hyperedge replacement grammar (HRG) is a context-free rewriting formal- ism for generating graphs (Drewes et al., 1997). Its want-01 believe-01 ARG1 ARG0 girl boy ARG0 ARG1 Figure 1: An example of AMR graph representing the meaning of: “The boy wants the girl to believe him” synchronous counterpart, SHRG, can be used for transforming a graph from/to another structured rep- resentation such as a string or tree structure. There- fore, an SHRG-based approach can be used for AMR parsing. Previous approaches usually first map the components of the sentence to components of the graph. Then different supervised algorithms are used to assemble these graph components to gen- erate a complete AMR graph (Flanigan et al., 2014; Wang et al., 2015b; Wang et al., 2015a). Previously, we have developed a system that learns SHRG rules from sentence/AMR graph pairs (Peng et al., 2015), with automatic alignments extracted from JAMR (Flanigan et al., 2014). Dur- ing the decoding procedure, we also use the concept identification results from Flanigan et al. (2014). The system is evaluated on the newswire section of LDC2013E117, which has around 4000 sentence- AMR pairs as training data. In this paper, we extend this system by using the alignments from Ulf Hermjakob’s automatic aligner and building a perceptron-based concept identifier where the boundary information of the mapped frag- 1185 want-01 believe-01 ARG1 ARG0 girl boy ARG0 ARG1 ARG0 X0 X2 X3 X1 X3 1 2 want-01 1 ARG1 2 X1 boy ARG0 ARG1 X2 boy X2 X3 X1 1 1 ARG0 ARG1 boy X3 X1 ARG0 X3 2 2 …

Content truncated to 500 words