Fast Boosting-based Part-of-Speech Tagging and Text Chunking with Efficient Rule Representation for Sequential Labeling

International Conference RANLP 2009 - Borovets, Bulgaria, pages 161–165 Fast Boosting-based Part-of-Speech Tagging and Text Chunking with Efficient Rule Representation for Sequential Labeling Tomoya Iwakura Fujitsu Laboratories Ltd. 1-1, Kamikodanaka 4-chome, Nakahara-ku, Kawasaki 211-8588, Japan iwakura.tomoya@jp.fujitsu.com Abstract This paper proposes two techniques for fast sequential labeling such as part-of-speech (POS) tagging and text chunking. The first technique is a boosting-based al- gorithm that learns rules represented by combination of features. To avoid time-consuming evaluation of combination, we divide features into not used ones and used ones for learning combination. The other is a rule representation. Usual POS taggers and text chunkers decide the tag of each word by using the features gen- erated from the word and its surrounding words. Thus similar rules, for example, that consist of the same set of words but only differ in locations from current words, are generated. We use a rule representation that enables us to merge such rules. We evaluate our meth- ods with POS tagging and text chunking. The experi- mental results show that our methods show faster pro- cessing speed than taggers and chunkers without our methods while maintaining accuracy. 1 Introduction Several machine learning algorithms such as Support Vec- tor Machines (SVMs) and boosting-based learning algo- rithms have been applied to Natural Language Processing (NLP) problems successfully. The cases of boosting in- clude text categorization [11], POS tagging [5] and text chunking [7, 5], and so on. Furthermore, parsers based on boosting-based learners have shown fast processing speed [7, 5]. However, to process large data such as WEB data and e-mails, processing speed of base technologies such as POS tagging and text chunking will be important. This paper proposes two techniques for improving pro- cessing speed of POS tagging and text chunking. The first technique is a boosting-based algorithm that learns rules. Instead of specifying combination of features manually, we specify features that are not used for the combination of features as atomic. Our boosting algorithm learns rules that consist of features or a feature from non-atomic features, and rules consisting of a feature from atomic features. The other is a rule representation for sequential label- ing such as POS tagging and text chunking. Usual POS taggers and text chunkers decide the tag of each word by using features generated from the current word and its sur- rounding words. Thus each word and its attributes, such as character-types, are evaluated several times in different relative locations from current word. We propose a repre- sentation that enables us to merge similar rules that consist of the same set of words and attributes that only differ in positions from current word. The experimental results with English POS tagging and text chunking show the taggers and chunkers based on our methods show faster processing speed than without our methods while maintaining competitive accuracy. 2 Boosting-based Learner 2.1 Preliminaries Let X be the set of examples and Y be a set of labels {−1, +1}. Let F = {f1, f2, ..., fM} be M …

Content truncated to 500 words