A Text Classifier Based on Sentence Category VSM

A Text Classifier Based on Sentence Category VSM ZHANG Yun-liang1, 2 ZHANG Quan1 1. Institute of Acoustics, Chinese Academy of Sciences, Beijing 100080, China 2. Graduate School of the Chinese Academy of Sciences, Beijing 100039, China zylioa@126.com, zhq@mail.ioa.ac.cn Abstract. VSM is a mature model of text representation for categorization. Words are commonly used as dimensions of feature space of VSM, but words only provide little semantic information. Sentence category theory is an important component of HNC theory and can provide abundant information about meaning, structure and style of a sentence. We use sentence categories as dimensions of feature space, reduce the dimensionality by dividing mixed sentence categories and reform the weights by tfc-weighting algorithm. By simple vector distance calculation, we can get the parameters of the classifier and execute the categorization. The average precision and recall of our classifier are acceptable and can be improved by other HNC techniques. Key words: Text categorization；HNC theory；Nature Language Processing; Sentence category； Computer application 1 Introduction With the rapid increase of electronic documents, especially the growth of the internet and intranet, automatic text processing becomes more and more important. Text categorization is usually the basic and important components of text processing applications. It will benefit other applications, such as IR, Q&A. Statistical categorization is the major method to resolve the categorization problem. Vector Space Model (VSM) was created by G. Salton in 1960s and now widely used in text representation [1]. All categorization methods based on VSM commonly include feature vector generation, dimensionality reduction of feature space, machine learning and categorization execution [2-4]. The idea of VSM representation is to discretize a consecutive text and form a vector in particular feature space. Words are the most common features of VSM and there are also models that take phrases, terms or Chinese character as features [5]. The weight of every element of a vector is usually the number of occurrences or its transform. Tfc-weighting is a transform algorithm. Typically, the dimensionality of the feature space is more than tens of thousands. It is too high to result in too much calculation time and memory space. So dimensionality reduction is necessary. Feature selection and feature merger are two useful approaches. Feature selection is to discard some features non-informative or non-dipartite. Document Frequency Threshold, Information Gain, CHI, Mutual Information are means of feature selection [6-10]. Feature merger is an approach to compress the feature space. Latent Semantic Indexing (LSI) and using Ontology such as WordNet, HowNet, thesaurus and HNC concept primitive woods are all feature merger [11-12]. With machine learning methods, we can get the parameters of the classifier and decide an incoming document belong to which category. Simple vector distance, naïve Bayes, k-Nearest Neighbor (k-NN), Support Vector Machine (SVM), Artificial Neural Network (ANN), Decision Tree, Voting and so on are all applied methods [13]. 244 2 Related HNC knowledge Chinese characters, words, phrases, terms are common features of VSM. With the increase of the feature granularity, the ambiguity decrease. But the dimension will increase because combination …

Content truncated to 500 words