A Joint Model for Parsing Syntactic and Semantic Dependencies

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 188–192 Manchester, August 2008 A Joint Model for Parsing Syntactic and Semantic Dependencies Xavier Llu´ıs and Llu´ıs M`arquez TALP Research Centre – Software Department (LSI) Technical University of Catalonia (UPC) {xlluis,lluism}@lsi.upc.edu Abstract This paper describes a system that jointly parses syntactic and semantic dependen- cies, presented at the CoNLL-2008 shared task (Surdeanu et al., 2008). It combines online Peceptron learning (Collins, 2002) with a parsing model based on the Eisner algorithm (Eisner, 1996), extended so as to jointly assign syntactic and semantic la- bels. Overall results are 78.11 global F1, 85.84 LAS, 70.35 semantic F1. Official re- sults for the shared task (63.29 global F1; 71.95 LAS; 54.52 semantic F1) were sig- nificantly lower due to bugs present at sub- mission time. 1 Introduction The main goal of this work was to construct a joint learning architecture for syntactic-semantic pars- ing and to test whether the syntactic and semantic layers can benefit each other from the global train- ing and inference. All the components of our system were built from scratch for this shared task. Due to strong time limitations, our design decisions were biased towards constructing a simple and feasible system. Our proposal is a first order linear model that re- lies on an online averaged Perceptron for learning (Collins, 2002) and an extended Eisner algorithm for the joint parsing inference. Systems based on Eisner algorithm (Carreras et al., 2006; Carreras, 2007) showed a competitive performance in the syntactic parsing of the English language in some past CoNLL shared tasks. Also, 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. we believe that extending the Eisner algorithm to jointly parse syntactic and semantic dependencies it is a natural step to follow. Note that syntactic and semantic tasks are re- lated but not identical. Semantic dependencies can take place between words loosely related by the syntactic structure. Another difficulty is that state of the art SRL systems (Surdeanu et al., 2007) strongly rely on features extracted from the syn- tactic tree. The joint model grows syntactic and semantic structures at the same time, so features extracted from the syntactic tree (e.g., a syntactic path between a modifier and a distant predicate) are not available or expensive to compute within the Eisner algorithm. We overcome this problem again with a very simple (though not elegant) solu- tion, consisting of introducing a previous syntactic parsing step. 2 System architecture This section briefly describes the main components of our system: 1) Preprocessing, 2) Syntactic pars- ing, 3) Predicate identification, 4) Joint syntactic- semantic parsing, and 5) Postprocessing. In preprocessing, the training corpus is traversed and feature extraction performed. Main features are borrowed from pre-existing well-known sys- tems (see next subsection). The initial syntactic parsing is based on an Eisner parser trained with Perceptron and it is merely intended to allow the extraction of syntactic-based features for all the following phases (which share …

Content truncated to 500 words