Enhanced Genre Classification through Linguistically Fine-Grained POS Tags

Enhanced Genre Classification through Linguistically Fine-Grained POS Tags∗ Alex Chengyu Fang1 and Jing Cao2 Department of Chinese, Translation and Linguistics City University of Hong Kong, Tat Chee Avenue, Kowloon, Hong Kong SAR 1acfang@cityu.edu.hk 2cjing3@student.cityu.edu.hk Abstract. We propose the use of fine-grained part-of-speech (POS) tags as discriminatory attributes for automatic genre classification and report empirical results from an experiment that indicate substantial accuracy gain by such features over the conventional bag-of-words approach through word unigrams. In particular, this paper reports our research to investigate the performance of a fine-grained tag set when tested with the British component of the International Corpus of English. Ten different genre classification tasks were identified and the performance of the tags was evaluated in terms of F-score. Our results show that the use of linguistically fine-grained POS tags produces superior accuracy when compared with word unigrams, particularly for a rich set of 32 different genres with Naïve Bayes Multinominal Classifier. Through a comparison with an impoverished tag set, our results further demonstrate that the superior performance is due to the rich linguistic information embodied in the 400-strong different POS tags. Keywords: automatic genre classification, ICE-GB, fine-grained POS tag, linguistic granularity, AUTASYS. ∗ The work described in this paper was supported partially by research grants (Nos 9610126, 7008002, 7002387 and 7002190) from City University of Hong Kong. The authors would like to thank the reviewers for their valuable comments and suggestions. Copyright 2010 by Alex Chengyu Fang and Jing Cao 1 Introduction Text classification has been conventionally based on content matters and sentiment polarities. There are situations where genre classification is required for the identification of, for example, formal and informal sources of information. Genre classification of text is a process of classifying texts or documents according to the criterion of genre, such as style, form, or purpose, based on the assumption that “a document can be represented by the values of features that seem to express the attribute of a genre” (Lim et al. 2005:1264). Part-of-speech (POS) tags have been employed in automatic genre classification in that they do not “reflect the topic of the document, but rather the type of text used in the document” (Finn and Kushmerick, 2003) and that their distribution has been observed to vary across different genres (e.g. Nakamura, 1993; Rayson et al., 2002). Nevertheless, a majority of past studies have included POS tags with other features to form a combined feature set. For example, Karlgren and Cutting (1994) included 6 POS tags (i.e. adverb, preposition, 2nd person pronoun, 1st person pronoun, noun and present verb) in classifying genres of the Brown Corpus. They carried out the classification tasks in terms of 2, 4 and 15 genre classes according to Brown categories. The combined feature set achieved an accuracy of 96%, 73% and 52% in the three classification tasks respectively. Dewdney et al. (2001) included POS tags of content words (i.e. noun, verb, adjective and adverb), where verbs were further defined in past, present and future tenses. Again, with a PACLIC …

Content truncated to 500 words