CMU: Arc-Factored, Discriminative Semantic Dependency Parsing

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 176–180, Dublin, Ireland, August 23-24, 2014. CMU: Arc-Factored, Discriminative Semantic Dependency Parsing Sam Thomson Brendan O’Connor Jeffrey Flanigan David Bamman Jesse Dodge Swabha Swayamdipta Nathan Schneider Chris Dyer Noah A. Smith Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {sthomson,brenocon,jflanigan,dbamman,jessed, swabha,nschneid,cdyer,nasmith}@cs.cmu.edu Abstract We present an arc-factored statistical model for semantic dependency parsing, as de- fined by the SemEval 2014 Shared Task 8 on Broad-Coverage Semantic Dependency Parsing. Our entry in the open track placed second in the competition. 1 Introduction The task of broad coverage semantic dependency parsing aims to provide a shallow semantic analysis of text not limited to a specific domain. As distinct from deeper semantic analysis (e.g., parsing to a full lambda-calculus logical form), shallow seman- tic parsing captures relationships between pairs of words or concepts in a sentence, and has wide application for information extraction, knowledge base population, and question answering (among others). We present here two systems that produce seman- tic dependency parses in the three formalisms of the SemEval 2014 Shared Task 8 on Broad-Coverage Semantic Dependency Parsing (Oepen et al., 2014). These systems generate parses by extracting fea- tures for each potential dependency arc and learn- ing a statistical model to discriminate between good arcs and bad; the first treats each labeled edge de- cision as an independent multiclass logistic regres- sion (§3.2.1), while the second predicts arcs as part of a graph-based structured support vector machine (§3.2.2). Common to both models is a rich set of features on arcs, described in §3.2.3. We include a discussion of features found to have no discernable effect, or negative effect, during development (§4). Our system placed second in the open track of the Broad-Coverage Semantic Dependency Parsing This work is licensed under a Creative Commons Attribution 4.0 International License. Page numbers and proceedings footer are added by the organizers. License details: http: //creativecommons.org/licenses/by/4.0/ Figure 1: Example annotations for DM (top), PAS (middle), and PCEDT (bottom). task (in which output from syntactic parsers and other outside resources can be used). We present our results in §5. 2 Formalisms The Shared Task 8 dataset consists of annota- tions of the WSJ Corpus in three different se- mantic dependency formalisms. DM is derived from LinGO English Resource Grammar (ERG) annotations in DeepBank (Flickinger et al., 2012). PAS is derived from the Enju HPSG treebank us- ing the conversion rules of Miyao et al. (2004). PCEDT is derived from the tectogrammatical layer of the Prague Czech-English Dependency Treebank (Hajiˇc, 1998). See Figure 1 for an example. The three formalisms come from very different linguistic theories, but all are represented as labeled directed graphs, with words as vertices, and all have “top” annotations, corresponding roughly to the semantic focus of the sentence. (A “top” need not be a root of the graph.) This allows us to use the same machinery (§3) for training and testing statistical models for the three formalisms. 3 Models We treat the …

Content truncated to 500 words