Voted Approach for Part of Speech Tagging in Bengali

Voted Approach for Part of Speech Tagging in Bengali* * The authors thankfully acknowledge the comments and suggestions of Anette Frank, Department of Computational Linguistics, University of Heidelberg, Germany, during preparation of the camera ready version. Copyright 2009 by Asif Ekbal, Md. Hasanuzzaman, and Sivaji Bandyopadhyay Asif Ekbal a, Md. Hasanuzzaman b, and Sivaji Bandyopadhyay c aDepartment of Computational Linguistics, University of Heidelberg, 1m Neuenheimer Feld 325, 69120 Heidelberg, Germany ekbal@cl.uni-heidelberg.de, asif.ekbal@gmail.com bWest Bengal Industrial Development Corporation, Kolkata, India hasanuzzaman.im@gmail.com cDepartment of Computer Science and Engineering, Jadavpur University, Kolkata-700032, India sivaji_cse_ju@yahoo.com, sbandyopadhyay@cse.jdvu.ac.in Abstract. Part of Speech (POS) tagging is the task of labeling each word in a sentence with its appropriate syntactic category called part of speech. POS tagging is a very important preprocessing task for language processing activities. In this paper, we report about our work on POS tagging for Bengali by combining different POS tagging systems using three weighted voting techniques. The individual POS taggers are based on Maximum Entropy (ME), Conditional Random Field (CRF) and Support Vector Machine (SVM) frameworks. The POS taggers use a tag set of 27 POS tags, defined for the Indian languages. The individual system makes use of the different contextual information of the words along with the variety of word-level features that are helpful in predicting the various POS classes. The POS tagger has been trained and tested with 57,341 and 35K tokens, respectively. It has been experimentally verified that the lexicon, named entity recognizer and different word suffixes are effective in handling the unknown word problems and improve the accuracy of the POS tagger significantly. Experimental results show the effectiveness of the proposed voted POS tagger with an accuracy of 92.35%, which is an improvement of 5.29% over the least performing ME based system and 2.23% over the best performing SVM based system. Keywords: Part of Speech (POS) tagging, Statistical techniques, Voting, Bengali. 1 Introduction Part of Speech (POS) tagging is the task of labeling each word in a sentence with its appropriate syntactic category called part of speech. POS tagging is a very important preprocessing task for various language processing activities. This helps in doing deep parsing of text and in developing information extraction systems, semantic processing etc. POS tagging for natural language texts are developed using linguistic rules, stochastic models or a combination of both. Stochastic models (Cutting et al., 1992; Merialdo, 1994; Brants, 2000) have been widely used in POS tagging task for simplicity and language independence of the models. Among stochastic models, Hidden Markov Models (HMMs) are quite popular. Development of a stochastic tagger requires large amount of annotated data. Stochastic taggers with more than 95% word-level accuracy have been developed for English, German and other European languages, for which large labeled data is available. The problem is difficult for Indian languages (ILs) due to the lack of such annotated large corpus. Simple HMMs do not work well when small amount of labeled data are used to estimate the model parameters. 120 23rd Pacific Asia Conference on …

Content truncated to 500 words