CAMR at SemEval-2016 Task 8: An Extended Transition-based AMR Parser

Proceedings of SemEval-2016, pages 1173–1178, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics CAMR at SemEval-2016 Task 8: An Extended Transition-based AMR Parser Chuan Wang Brandeis University cwang24@brandeis.edu Sameer Pradhan cemantix.org Boulder Learning Inc. pradhan@cemantix.org Nianwen Xue Brandeis University xuen@brandeis.edu Xiaoman Pan Rensselaer Polytechnic Institute panx2@rpi.edu Heng Ji Rensselaer Polytechnic Institute jih@rpi.edu Abstract This paper describes CAMR, the transition- based parser that we use in the SemEval-2016 Meaning Representation Parsing task. The main contribution of this paper is a descrip- tion of the additional sources of information that we use as features in the parsing model to further boost its performance. We start with our existing AMR parser and experiment with three sets of new features: 1) rich named en- tities, 2) a verbalization list, 3) semantic role labels. We also use the RPI Wikifier to wikify the concepts in the AMR graph. Our parser achieves a Smatch F-score of 62% on the offi- cial blind test set. 1 Introduction AMR parsing is the task of taking a sentence as input and producing as output an Abstract Mean- ing Representation (AMR) that is a rooted, di- rected, edge-labeled and leaf-labeled graph that is used to represent the meaning of a sentence (Ba- narescu et al., 2013). AMR parsing has drawn an increasing amount of attention recently. The first published AMR parser, JAMR (Flanigan et al., 2014), performs AMR parsing in two stages: con- cept identification and relation identification. Flani- gan et al. (2014) treat concept identification as a sequence labeling task and utilize a semi-Markov model to map spans of words in a sentence to concept graph fragments. For relation identifica- tion, they adopt graph-based techniques similar to those used in dependency parsing (McDonald et al., 2005). Instead of finding maximum spanning trees (MST) over words, they propose an algorithm that finds the maximum spanning connected subgraph (MSCG) over concept fragments identified in the first stage. Wang et al. (2015b) describes a transition-based parser that also involves two stages. In the first step, an input sentence is parsed into a dependency tree with a dependency parser. In the second step, it transforms the dependency tree into an AMR graph by performing a series of actions. Note that the de- pendency parser used in the first step can be any off-the-shelf dependency parser and does not have to trained on the same data set as used in the second step. There are also approaches which utilize gram- mar induction to parse the AMR. Artzi et al. (2015) presents a model that first use Combinatory Cat- egorial Grammar (CCG) to construct the lambda- calculus representations of the sentence, then further resolve non-compositional dependencies using a factor graph. Peng et al.(2015) and Pust et al.(2015) formalize parsing AMR as a machine translation problem by learning string-graph/string-tree rules from the annotated data. Although the field of AMR parsing is growing and several systems (Wang et al., 2015a; Artzi et al., 2015; Pust et al., 2015; Flanigan et al., 2014) have substantially …

Content truncated to 500 words