Semantic Role Labeling via Consensus in Pattern-Matching

Proceedings of the 9th Conference on Computational Natural Language Learning (CoNLL), pages 185–188, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Semantic Role Labeling via Consensus in Pattern-Matching Chi-San (Althon) Lin Tony C. Smith Department of Computer Science Department of Computer Science Waikato University Waikato University Hamilton, New Zealand Hamilton, New Zealand cl123@cs.waikato.ac.nz tcs@cs.waikato.ac.nz Abstract This paper describes a system for semantic role labeling for the CoNLL2005 Shared task. We divide the task into two sub-tasks: boundary recognition by a general tree- based predicate-argument recognition algo- rithm to convert a parse tree into a flat rep- resentation of all predicates and their related boundaries, and role labeling by a consensus model using a pattern-matching framework to find suitable roles for core constituents and adjuncts. We describe the system architecture and report results for the CoNLL2005 development dataset. 1 Introduction Semantic role labeling is to find all arguments for all predicates in a sentence, and classify them by semantic roles such as A0, A1, AM-TMP and so on. The performance of semantic role labeling can play a key role in Natural Language Processing applications, such as Information Extraction, Ques- tion Answering, and Summarization (Pradhan et al., 2004). Most existing systems separate semantic role la- beling into two sub-problems, boundary recogni- tion and role classification, and use feature-based models to address both (Carreras et al., 2004). Our strategy is to develop a boundary analyzer by a general tree-based predicate-argument recognition algorithm (GT-PARA) for boundary recognition, and a pattern-matching model for role classifica- tion. The only information used in our system is Charniak’s annotation with words, which contains all useful syntactic annotations. Five features, which are Headword, Phrase type, Voice, Target verb, and Preposition (of the first word), and a Pat- tern set, which includes numbers and types of roles in a pattern, are used for the pattern-matching ap- proach. We develop a Pattern Database, trained by Wall Street Journal section 02 to 21, as our knowl- edge/Data base. The system outline is described in the following section. 2 System Description An overview of the system architecture is shown in Figure 1. The input is a full parse tree for each sentence. We convert a sentence with words, and Charniak’s information into a parsed tree as the input of GT-PARA. GT-PARA then converts the parse tree into a flat representation with all predi- cates and arguments expressed in [GPLVR] for- mat; where G: Grammatical function – 5 denotes subject, 3 object, and 2 others; P: Phrase type of this boundary – 00 denotes ADJP, 01 ADVP, 02 NP, 03 PP, 04 S, 05 SBAR, 06 SBARQ, 07 SINV, 08 SQ, 09 VP, 10 WHADVP, 11 WHNP, 12 WHPP, and 13 Others L: Distance (and position) of the argument with respect to the predicate that follows V: Voice of the predicate, 0: active 1: passive R: Distance (and position) of the argument with respect to the preceding predicate (n.b. L and R are mutually exclusive). An example of the output of GT-PARA is …

Content truncated to 500 words