Adjective Density as a Text Formality Characteristic for Automatic Text Classification: A Study Based on the British National Corpus

Adjective Density as a Text Formality Characteristic for Automatic Text Classification: A Study Based on the British National Corpus∗∗∗∗ Alex Chengyu Fang and Jing Cao Department of Chinese, Translation and Linguistics City University of Hong Kong, Tat Chee Avenue, Kowloon, Hong Kong SAR acfang@cityu.edu.hk, cjing3@student.cityu.edu.hk Abstract. In this article, we report significant findings resulting from an investigation into the correlation between adjective density, calculated as the proportion of adjectives in word tokens, and degrees of text formality as part of an attempt to examine the potential application of adjectives in automatic text classification and identification. Correlations obtained from the training corpus will be compared with human ranking of the text categories concerned in the study and then adapted to unseen data in the test set. A linear regression analysis suggests a strong correlation between degrees of text formality and adjective density. With a weighted average F-measure of 0.606 achieved by a Naïve Bayes classifier, the research establishes adjectives as a powerful differentia of text categories amongst the open word classes, an important feature that has been generally ignored by past studies in automatic text categorization. The empirical findings suggest that the use of adjective density will lead to enhanced practical systems for automatic text classification. Keywords: Adjective density, text formality, text classification, Linear Regression, Naïve Bayes. ∗ The work described in this paper was supported partially by research grants (Nos 7002190, 7200120 and 7002387) from City University of Hong Kong. The authors would like to thank the reviewers for their valuable comments and suggestions. Copyright 2009 by Alex Chengyu Fang and Jing Cao 1 Introduction 1.1 Text Classification and Past Studies Text classification according to degrees of formality has been a long-standing issue that is both linguistically complex and computationally challenging. Biber (1988) examines 23 text categories according to six dimensions involving 67 chosen linguistic features. The first dimension (involved vs. informational) is related to text formality although the study reveals no clear-cut distinction between spoken and written texts through the use of the selected linguistic features. Sigley (1997) proposes to evaluate text formality by integrating 29 linguistic features into a formality index. Although the study concludes that such an index can evaluate a wide variety of text categories, there are still overlapping written and spoken categories that the formality index cannot deal with satisfactorily. More recently, linguistic features specifically involving word classes have been investigated and employed for automatic text classification either through combining various word classes into a joint index to measure text formality or through using specific sets of a certain grammatical class to distinguish text categories. For example, Heylighen and Dewaele (1999 and 2002) propose a formula of formality based on word classes: 130 23rd Pacific Asia Conference on Language, Information and Computation, pages 130–139 F = freq(noun,adjective, preposition,article) − ∑freq(pronoun, verb,adverb, interjection) +100 ∑ 2 (1) Results show that such a formality score is quite effective in separating speech from writing and that imaginative writing can be separated from informational writing. Dempsey et al. (2007) …

Content truncated to 500 words