Semantic Role Labeling via Instance-Based Learning

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 180–188, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Semantic Role Labeling via Instance-Based Learning Chi-San Althon Lin Department of Computer Science Waikato University Hamilton, New Zealand cl123@cs.waikato.ac.nz Tony C. Smith Department of Computer Science Waikato University Hamilton, New Zealand tcs@cs.waikato.ac.nz Abstract This paper demonstrates two methods to improve the performance of instance- based learning (IBL) algorithms for the problem of Semantic Role Labeling (SRL). Two IBL algorithms are utilized: k-Nearest Neighbor (kNN), and Priority Maximum Likelihood (PML) with a modified back-off combination method. The experimental data are the WSJ23 and Brown Corpus test sets from the CoNLL- 2005 Shared Task. It is shown that ap- plying the Tree-Based Predicate- Argument Recognition Algorithm (PARA) to the data as a preprocessing stage allows kNN and PML to deliver F1: 68.61 and 71.02 respectively on the WSJ23, and F1: 56.96 and 60.55 on the Brown Corpus; an increase of 8.28 in F1 measurement over the most recent pub- lished PML results for this problem (Palmer et al., 2005). Training times for IBL algorithms are very much faster than for other widely used techniques for SRL (e.g. parsing, support vector machines, perceptrons, etc); and the feature reduc- tion effects of PARA yield testing and processing speeds of around 1.0 second per sentence for kNN and 0.9 second per sentence for PML respectively, suggest- ing that IBL could be a more practical way to perform SRL for NLP applica- tions where it is employed; such as real- time Machine Translation or Automatic Speech Recognition. 1 Introduction The proceedings from CoNLL2004 and CoNLL2005 detail a wide variety of approaches to Semantic Role Labeling (SRL). Many re- search efforts utilize machine learning (ML) ap- proaches; such as support vector machines (Mo- schitti et al., 2004; Pradhan et al., 2004), percep- trons (Carreras et al., 2004), the SNoW learning architecture (Punyakanok et al., 2004), EM- based clustering (Baldewein et al., 2004), trans- formation-based learning (Higgins, 2004), mem- ory-based learning (Kouchnir, 2004), and induc- tive learning (Surdeanu et al., 2003). This paper compares two instance-based learning ap- proaches, kNN and PML. The PML method used here utilizes a modification of the backoff lattice method used by Gildea & Jurafsky (2002) to use a set of basic features—specifically, the features employed for learning in this paper are Predicate (pr), Voice (vo), Phrase Type (pt), Dis- tance (di), Head Word (hw), Path (pa), Preposi- tion in a PP (pp), and an “Actor” heuristic. The general approach presented here is an example of memory-based learning. Many existing SRL systems are also memory-based (Bosch et al., 2004;Kouchnir, 2004), implemented using TilMBL software (http://ilk.kub.nl/software.html) with advanced methods such as Feature Weighting, and so forth. This paper measures the performance of kNN and PML for comparison in terms of accuracy and processing speed, both against each other and against previously published results. 2 Related Work Features Most of the systems outlined in CoNLL2004 and CoNLL2005 utilize as many as …

Content truncated to 500 words