Exploiting Full Parsing Information to Label Semantic Roles Using an Ensemble of ME and SVM via Integer Linear Programming

Proceedings of the 9th Conference on Computational Natural Language Learning (CoNLL), pages 233–236, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Exploiting Full Parsing Information to Label Semantic Roles Using an Ensemble of ME and SVM via Integer Linear Programming Tzong-Han Tsai, Chia-Wei Wu, Yu-Chun Lin, Wen-Lian Hsu Institute of Information Science Academia Sinica Taipei 115, Taiwan {thtsai, cwwu, sbb, hsu}@iis.sinica.edu.tw Abstract In this paper, we propose a method that exploits full parsing information by repre- senting it as features of argument classifi- cation models and as constraints in integer linear learning programs. In addition, to take advantage of SVM-based and Maxi- mum Entropy-based argument classifica- tion models, we incorporate their scoring matrices, and use the combined matrix in the above-mentioned integer linear pro- grams. The experimental results show that full parsing information not only in- creases the F-score of argument classifi- cation models by 0.7%, but also effectively removes all labeling inconsis- tencies, which increases the F-score by 0.64%. The ensemble of SVM and ME also boosts the F-score by 0.77%. Our system achieves an F-score of 76.53% in the development set and 76.38% in Test WSJ. 1 Introduction The Semantic Role Labeling problem can be for- mulated as a sentence tagging problem. A sentence can be represented as a sequence of words, as phrases (chunks), or as a parsing tree. The basic units of a sentence are words, phrases, and con- stituents in these representations, respectively.. Pradhan et al. (2004) established that Constituent- by-Constituent (C-by-C) is better than Phrase-by- Phrase (P-by-P), which is better than Word-by- Word (W-by-W). This is probably because the boundaries of the constituents coincide with the arguments; therefore, C-by-C has the highest ar- gument identification F-score among the three ap- proaches. In addition, a full parsing tree also provides richer syntactic information than a sequence of chunks or words. Pradhan et al. (2004) compared the seven most common features as well as several features related to the target constituent’s parent and sibling constituents. Their experimental results show that using other constituents’ information increases the F-score by 6%. Punyakanok et al. (2004) represent full parsing information as con- straints in integer linear programs. Their experi- mental results show that using such information increases the argument classification accuracy by 1%. In this paper, we not only add more full parsing features to argument classification models, but also represent full parsing information as constraints in integer linear programs (ILP) to resolve label in- consistencies. We also build an ensemble of two argument classification models: Maximum Entropy and SVM by combining their argument classifica- tion results and applying them to the above- mentioned ILPs. 2 System Architecture Our SRL system is comprised of four stages: prun- ing, argument classification, classification model incorporation, and integer linear programming. This section describes how we build these stages, including the features used in training the argu- ment classification models. 2.1 Pruning 233 When the full parsing tree of a sentence is avail- able, only the constituents in the tree are consid- …

Content truncated to 500 words