Mixing and Blending Syntactic and Semantic Dependencies

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 248–252 Manchester, August 2008 Mixing and Blending Syntactic and Semantic Dependencies Yvonne Samuelsson Dept. of Linguistics Stockholm University yvonne.samuelsson@ling.su.se Johan Eklund SSLIS University College of Bor˚as johan.eklund@hb.se Oscar T¨ackstr¨om Dept. of Linguistics and Philology SICS / Uppsala University oscar@sics.se Mark Fiˇsel Dept. of Computer Science University of Tartu fishel@ut.ee Sumithra Velupillai Dept. of Computer and Systems Sciences Stockholm University / KTH sumithra@dsv.su.se Markus Saers Dept. of Linguistics and Philology Uppsala University markus.saers@lingfil.uu.se Abstract Our system for the CoNLL 2008 shared task uses a set of individual parsers, a set of stand-alone semantic role labellers, and a joint system for parsing and semantic role labelling, all blended together. The system achieved a macro averaged labelled F1- score of 79.79 (WSJ 80.92, Brown 70.49) for the overall task. The labelled attach- ment score for syntactic dependencies was 86.63 (WSJ 87.36, Brown 80.77) and the labelled F1-score for semantic dependen- cies was 72.94 (WSJ 74.47, Brown 60.18). 1 Introduction This paper presents a system for the CoNLL 2008 shared task on joint learning of syntactic and se- mantic dependencies (Surdeanu et al., 2008), com- bining a two-step pipelined approach with a joint approach. In the pipelined system, eight different syntac- tic parses were blended, yielding the input for two variants of a semantic role labelling (SRL) system. Furthermore, one of the syntactic parses was used with an early version of the SRL system, to pro- vide predicate predictions for a joint syntactic and semantic parser. For the final submission, all nine syntactic parses and all three semantic parses were blended. The system is outlined in Figure 1; the dashed arrow indicates the potential for using the predi- c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. Process Parse + SRL Process 8 MaltParsers Parser Blender Process 2 Pipelined SRLs SRL Blender Joint Parser/SRL Possible Iteration Figure 1: Overview of the submitted system. cate prediction to improve the joint syntactic and semantic system. 2 Dependency Parsing The initial parsing system was created using Malt- Parser (Nivre et al., 2007) by blending eight dif- ferent parsers. To further advance the syntactic ac- curacy, we added the syntactic structure predicted by a joint system for syntactic and semantic depen- dencies (see Section 3.4) in the blending process. 2.1 Parsers The MaltParser is a dependency parser genera- tor, with three parsing algorithms: Nivre’s arc standard, Nivre’s arc eager (see Nivre (2004) for a comparison between the two Nivre algo- rithms), and Covington’s (Covington, 2001). Both of Nivre’s algorithms assume projectivity, but the MaltParser supports pseudo-projective parsing (Nilsson et al., 2007), for projectivization and de- projectivization. 248 WSJ Brown Best single parse 85.22% 78.37% LAS weights 87.00% 80.60% Learned weights 87.36% 80.77% Table 1: Labelled attachment score on the two test sets of the best single parse, blended with weights set to PoS labelled attachment score (LAS) and blended with learned weights. Four parsing algorithms (the two …

Content truncated to 500 words