Improving Chunk-based Semantic Role Labeling with Lexical Features

Proceedings of Recent Advances in Natural Language Processing, pages 226–232, Hissar, Bulgaria, 12-14 September 2011. Improving Chunk-based Semantic Role Labeling with Lexical Features Wilker Aziz, Miguel Rios and Lucia Specia Research Group in Computational Linguistics University of Wolverhampton Stafford Street, Wolverhampton, WV1 1SB, UK {w.aziz, m.rios, l.specia}@wlv.ac.uk Abstract We present an approach for Seman- tic Role Labeling (SRL) using Condi- tional Random Fields in a joint identifi- cation/classification step. The approach is based on shallow syntactic information (chunks) and a number of lexicalized fea- tures such as selectional preferences and automatically inferred similar words, ex- tracted using lexical databases and distri- butional similarity metrics. We use se- mantic annotations from the Proposition Bank for training and evaluate the system using CoNLL-2005 test sets. The addi- tional lexical information led to improve- ments of 15% (in-domain evaluation) and 12% (out-of-domain evaluation) on over- all semantic role classification in terms of F-measure. The gains come mostly from a better recall, which suggests that the addi- tion of richer lexical information can im- prove the coverage of existing SRL mod- els even when very little syntactic knowl- edge is available. 1 Introduction Identifying the relations that words or groups of words have with verbs in a sentence constitutes an important step for many applications in Natural Language Processing (NLP). This is addressed by the field of Semantic Role Labeling (SRL). SRL has been shown to contribute to many NLP appli- cations, such as Information Extraction, Question Answering and Machine Translation. Most of the SRL approaches operate via two consecutive steps: i) the identification of the argu- ments of a target predicate and ii) the classification of those arguments (Gildea and Jurafsky, 2002; Pradhan et al., 2004). Alternatively, graph mod- els can rely on the sequential nature of the shallow semantic parsing and perform both SRL steps si- multaneously (Roth and tau Yih, 2005; Cohn and Blunsom, 2005). Features for SRL are usually extracted from chunks or constituent parse trees. While parse trees allow a set of very informative path-based, structural features, chunks can provide more re- liable annotations. Hacioglu et al. (2004) pro- pose the use of base phrases as data representation using Support Vector Machines in order to per- form a single argument classification step. Roth and tau Yih (2005) use the same sort of repre- sentation with Conditional Random Fields (CRF) as learning algorithm, motivated by the sequen- tial nature of the task. Cohn and Blunsom (2005) use CRF to perform SRL in a single identifica- tion/classification step based on features from con- stituent trees. Pradhan et al. (2008) point out the lack of se- mantic features as the bottleneck in argument role classification, a task closely-related to that of word sense disambiguation. Shallow lexical features such as word forms and word lemmas are very sparse. Although named-entity categories have been proposed to alleviate this sparsity problem, they only apply to a fraction of the arguments’ words. In this paper we propose the addition of other forms of lexical knowledge in …

Content truncated to 500 words