The Effect of Temporal-based Term Selection for Text Classification

The Effect of Temporal-based Term Selection for Text Classification Fumiyo Fukumoto1, Shogo Ushiyama2, Yoshimi Suzuki1 and Suguru Matsuyoshi1 1Interdisciplinary Graduate School of Medicine and Engineering, University of Yamanashi 2Faculty of Engineering, University of Yamanashi Kofu, 400-8511, JAPAN {fukumoto, t08kg006, ysuzuki, sugurum}@yamanashi.ac.jp Abstract This paper addresses the text classification problem that training data may derive from a different time period from the test data. We present a method of temporal-based term selection for timeline adaptation. We selected two types of informative terms ac- cording to corpus statistics. One is tempo- ral independent terms that are salient re- gardless of the timeline. Another is tem- poral dependent terms which are impor- tant for a specific period of time. For tem- poral dependent terms extracted from the training documents, we applied weight- ing function that weights terms accord- ing to the temporal distance between train- ing and test data in the process of train- ing classifiers. The results using Mainichi Japanese newspaper documents showed improvement over the three baselines. 1 Introduction Text classification supports and improves several tasks such as automated topic tagging, building topic directory, spam filtering, creating digital li- braries, sentiment analysis in user reviews, Infor- mation Retrieval, and even helping users to inter- act with search engines (Mourao et al., 2008). A growing number of machine learning techniques have been applied to text classification (Xue et al., 2008; Gopal and Yang, 2010). The common ap- proach is the use of term selection. Each document is represented using a vector of selected terms (Yang and Pedersen, 1997; Hassan et al., 2007). Then, they used training documents with category label to train classifiers. Once category models are trained, each document of the test data is classi- fied by using these models. Terms in the docu- ments may be considered more important to build the classification model according to the timelines, while the majority of supervised classification methods consider that each term provides equally information regardless to a period. For instance, as shown in Figure 1, the term “earthquake” ap- peared more frequently in the category “Science” than “International” early in 1995. However, it ap- peared frequently in the category “International” than “Science” since Sumatra earthquake occurred just off the southern coast of Sumatra, Indonesia in 2005. Similarly, the term “Alcindo” frequently ap- peared in the documents tagged “Sports” in 1994, since Alcindo is a Brazilian soccer player and he was one of the most loved players in 1994. The term did not appear more frequently in the “Sports” category since he retired in 1997. These observations show that salient terms in the training data, are not salient in the test data when training data may derive from a different time period from the test data. In this paper, we present a method for text clas- sification concerned with the impact that the varia- tion of the strength of term-class relationship over time. We selected two types of informative terms according to corpus statistics. One is temporal in- dependent terms that are salient …

Content truncated to 500 words