HIT-IR-WSD: A WSD System for English Lexical Sample Task

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 165–168, Prague, June 2007. c⃝2007 Association for Computational Linguistics HIT-IR-WSD: A WSD System for English Lexical Sample Task Yuhang Guo, Wanxiang Che, Yuxuan Hu, Wei Zhang and Ting Liu Information Retrieval Lab Harbin Institute of technology Harbin, China, 150001 {yhguo,wxche}@ir.hit.edu.cn Abstract HIT-IR-WSD is a word sense disambigua- tion (WSD) system developed for English lexical sample task (Task 11) of Semeval 2007 by Information Retrieval Lab, Harbin Institute of Technology. The system is based on a supervised method using an SVM classifier. Multi-resources including words in the surrounding context, the part- of-speech of neighboring words, colloca- tions and syntactic relations are used. The final micro-avg raw score achieves 81.9% on the test set, the best one among partici- pating runs. 1 Introduction Lexical sample task is a kind of WSD evaluation task providing training and test data in which a small pre-selected set of target words is chosen and the target words are marked up. In the training data the target words’ senses are given, but in the test data are not and need to be predicted by task par- ticipants. HIT-IR-WSD regards the lexical sample task as a classification problem, and devotes to extract effective features from the instances. We didn’t use any additional training data besides the official ones the task organizers provided. Section 2 gives the architecture of this system. As the task pro- vides correct word sense for each instance, a su- pervised learning approach is used. In this system, we choose Support Vector Machine (SVM) as classifier. SVM is introduced in section 3. Know- ledge sources are presented in section 4. The last section discusses the experimental results and present the main conclusion of the work performed. 2 The Architecture of the System HIT-IR-WSD system consists of 2 parts: feature extraction and classification. Figure 1 portrays the architecture of the system. Figure 1: The architecture of HIT‐IR‐WSD 165 Features are extracted from original instances and are made into digitized features to feed the SVM classifier. The classifier gets the features of training data to make a model of the target word. Then it uses the model to predict the sense of target word in the test data. 3 Learning Algorithm SVM is an effective learning algorithm to WSD (Lee and Ng, 2002). The SVM tries to find a hyperplane with the largest margin separating the training samples into two classes. The instances in the same side of the hyperplane have the same class label. A test instance’s feature decides the position where the sample is in the feature space and which side of the hyperplane it is. In this way, it leads to get a prediction. SVM could be extended to tackle multi-classes problems by using one- against-one or one-against-rest strategy. In the WSD problem, input of SVM is the fea- ture vector of the instance. Features that appear in all the training samples are arranged as a vector space. Every instance is mapped to …

Content truncated to 500 words