Negation Naive Bayes for Categorization of Product Pages on the Web

Proceedings of Recent Advances in Natural Language Processing, pages 586–591, Hissar, Bulgaria, 12-14 September 2011. Negation Naive Bayes for Categorization of Product Pages on the Web Kanako Komiya 1 Naoto Sato 1 Koji Fujimoto 1,2 Yoshiyuki Kotani 1 Tokyo University of Agriculture and Technology 1 Tensor Consulting Co.Ltd. 2 {kkomiya, kotani}@cc.tuat.ac.jp 50009646113@st.tuat.ac.jp koji.fujimoto@tensor.co.jp Abstract We propose the negation naive Bayes (NNB): a new method to categorize prod- uct pages on the Web depending on their information. It is a modified version of the naive Bayes (NB) and we got the idea from the complement naive Bayes (CNB). We compared the NNB with the NB and the CNB. Our experiments show that the NNB outperformed the other methods sig- nificantly when the product pages were distributed non-uniformly through cate- gories. 1 Introduction In late years, e-commerce, the services by which users can easily purchase products on the Web without visiting a store, is introduced in many companies. When products are purchased via In- ternet, the user narrows down the candidate cat- egories of each product in incremental steps. We categorized the products on the Web automatically depending on their information using the method of text classification (Sato et al., 2011). Many researchers have investigated text classifi- cation and the naive Bayes (NB) is one of the most famous methods for it. However when we use the NB classifier to categorize the products, the accu- racies were not very high, especially when the data distribution is very skewed. Hence this paper proposes the negation naive Bayes (NNB): a new method of text classifica- tion especially for the product pages on the Web depending on their information. It is a modified version of the NB and we got the idea from the complement naive Bayes (CNB). Our experiments showed that the NNB outperformed the NB and the CNB when the product pages were distributed non-uniformly. This paper is organized as follows. Section 2 re- views related works on text classifications and the NB classifiers. Section 3 describes the classifica- tion methods including our proposal method: the NNB. Section 4 describes the system to categorize the product pages and explains the experimental setting. We describe results in Section 5 and dis- cuss them in Section 6. Finally, we conclude the paper in Section 7. 2 Related Work Many works on text classification have been ac- complished so far. Approaches of Bayes are often used within the area of text classification (Mochi- hashi, 2006). Izutsu et al. (2005) categorized the html documents and compared the NB classi- fier with discriminant analysis and the rule-based method. They suggested the simple implemen- tation and the high scalability of the NB classi- fier. McCallum and Nigam (1998) suggested the difference between multinomial model and multi- variate Bernoulli model of the NB classifier in text classification. Lewis (1992) compared the differ- ence of the effect between the types of features used for text classification: words, phrases, clus- tered words, clustered phrases and indexing terms. W.Church (2000) used a concept …

Content truncated to 500 words