Semantic Role Labeling System Using Maximum Entropy Classifier

Proceedings of the 9th Conference on Computational Natural Language Learning (CoNLL), pages 189–192, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Semantic Role Lableing System using Maximum Entropy Classifier ∗ Ting Liu, Wanxiang Che, Sheng Li, Yuxuan Hu and Huaijun Liu Information Retrieval Lab School of Computer Science and Technology Harbin Institute of Technology China, 150001 {tliu, car, ls, yxhu, hjliu}@ir.hit.edu.cn Abstract A maximum entropy classifier is used in our semantic role labeling system, which takes syntactic constituents as the labeling units. The maximum entropy classifier is trained to identify and classify the predi- cates’ semantic arguments together. Only the constituents with the largest probabil- ity among embedding ones are kept. Af- ter predicting all arguments which have matching constituents in full parsing trees, a simple rule-based post-processing is ap- plied to correct the arguments which have no matching constituents in these trees. Some useful features and their combina- tions are evaluated. 1 Introduction The semantic role labeling (SRL) is to assign syn- tactic constituents with semantic roles (arguments) of predicates (most frequently verbs) in sentences. A semantic role is the relationship that a syntactic constituent has with a predicate. Typical semantic arguments include Agent, Patient, Instrument, etc. and also adjunctive arguments indicating Locative, Temporal, Manner, Cause, etc. It can be used in lots of natural language processing application sys- tems in which some kind of semantic interpretation is needed, such as question and answering, informa- tion extraction, machine translation, paraphrasing, and so on. ∗This research was supported by National Natural Science Foundation of China via grant 60435020 Last year, CoNLL-2004 hold a semantic role la- beling shared task (Carreras and M`arquez, 2004) to test the participant systems’ performance based on shallow syntactic parser results. In 2005, SRL shared task is continued (Carreras and M`arquez, 2005), because it is a complex task and now it is far from desired performance. In our SRL system, we select maximum en- tropy (Berger et al., 1996) as a classifier to im- plement the semantic role labeling system. Dif- ferent from the best classifier reported in litera- tures (Pradhan et al., 2005) – support vector ma- chines (SVMs) (Vapnik, 1995), it is much eas- ier for maximum entropy classifier to handle the multi-class classification problem without additional post-processing steps. The classifier is much faster than training SVMs classifiers. In addition, max- imum entropy classifier can be tuned to minimize over-fitting by adjusting gaussian prior. Xue and Palmer (2004; 2005) and Kwon et al. (2004) have applied the maximum entropy classifier to semantic role labeling task successfully. In the following sections, we will describe our system and report our results on development and test sets. 2 System Description 2.1 Constituent-by-Constituent We use syntactic constituent as the unit of labeling. However, it is impossible for each argument to find its matching constituent in all auto parsing trees. Ac- cording to statistics, about 10% arguments have no matching constituents in the training set of 245,353 189 constituents. The top five arguments with no match- ing constituents …

Content truncated to 500 words