Potsdam: Semantic Dependency Parsing by Bidirectional Graph-Tree Transformations and Syntactic Parsing

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 465–470, Dublin, Ireland, August 23-24, 2014. Potsdam: Semantic Dependency Parsing by Bidirectional Graph-Tree Transformations and Syntactic Parsing ˇZeljko Agi´c University of Potsdam zagic@uni-potsdam.de Alexander Koller University of Potsdam koller@ling.uni-potsdam.de Abstract We present the Potsdam systems that par- ticipated in the semantic dependency pars- ing shared task of SemEval 2014. They are based on linguistically motivated bidi- rectional transformations between graphs and trees and on utilization of syntactic de- pendency parsing. They were entered in both the closed track and the open track of the challenge, recording a peak average labeled F1 score of 78.60. 1 Introduction In the semantic dependency parsing (SDP) task of SemEval 2014, the meaning of a sentence is repre- sented in terms of binary head-argument relations between the lexical units – bi-lexical dependencies (Oepen et al., 2014). Since words can be seman- tic dependents of multiple other words, this frame- work results in graph representations of sentence meaning. For the SDP task, three such annotation layers are provided on top of the WSJ text of the Penn Treebank (PTB) (Marcus et al., 1993): – DM: the reduction of DeepBank HPSG anno- tation (Flickinger et al., 2012) into bi-lexical dependencies following (Oepen and Lønning, 2006; Ivanova et al., 2012), – PAS: the predicate-argument structures derived from the training set of the Enju HPSG parser (Miyao et al., 2004) and – PCEDT: a subset of the tectogrammatical anno- tation layer from the English side of the Prague Czech-English Dependency Treebank (Cinkov´a et al., 2009). The three annotation schemes provide three di- rected graph representations for each PTB sen- This work is licenced 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/ tence, with word forms as nodes and labeled de- pendency relations as edges pointing from func- tors to arguments. The SDP-annotated PTB text is split into training (sections 00–19), development (sec. 20) and testing sets (sec. 21). This in turn makes the SDP parsing task a problem of data- driven graph parsing, in which systems are to be trained for producing dependency graph represen- tations of sentences respecting the three underly- ing schemes. While a number of theoretical and preliminary contributions to data-driven graph parsing exist (Sagae and Tsujii, 2008; Das et al., 2010; Jones et al., 2013; Chiang et al., 2013; Henderson et al., 2013), our goal here is to investigate the sim- plest approach that can achieve competitive per- formance. Our starting point is the observation that the SDP graphs are relatively tree-like. On it, we build a system for data-driven graph parsing by (1) transforming dependency graphs into depen- dency trees in preprocessing, (2) training and us- ing syntactic dependency parsers over these trees and (3) transforming their output back into graphs in postprocessing. This way, we inherit the accu- racy and speed of syntactic dependency parsers. The secondary benefit is insight into the struc- ture of the semantic …

Content truncated to 500 words