Automatic Corpora Construction for Text Classification

International Joint Conference on Natural Language Processing, pages 726–732, Nagoya, Japan, 14-18 October 2013. Automatic Corpora Construction for Text Classification Dandan Wang, Qingcai Chen, Xiaolong Wang, Bingyang Yu Key Laboratory of Network Oriented Intelligent Computation Harbin Institute of Technology Shenzhen Graduate School, China {wangdandanhit, qingcai.chen}@gmail.com wangxl@insun.hit.edu.cn bingyang.yu@hotmail.com Abstract Since the machines become more and more intelligent, it is reasonable to expect the automatic construction of text classi- fiers by given just the objective categories. As trade-off solutions, existing researches usually provide additional information to the category terms to enhance the perfor- mance of a classifier. Unique from them, in this paper, we construct the standard corpora from the web by just providing text categories. Since there are million- s of manually constructed websites, it is hopeful to find out proper text categoriza- tion (TC) knowledge. So we directly go to the web and use the hierarchies implied in navigation bars to extract and verify TC re- sources. By addressing the issues of nav- igation bar recognition and text filtering, the corpora are constructed for given tex- t categories and the classifiers are trained based on them. We conduct our experi- ments on the large scale of webpages col- lected from the 500 top English websites on Alexa. The Open Directory Project (ODP) is used as testing corpus. Experi- mental results show that, being compared with the classifier based on manually la- beled corpus, the classifier trained on au- to constructed corpora reaches compara- ble performance for the categories that are well covered by the training corpus. 1 Introduction As one of the key techniques in web information processing, text classification has been studied for a long time (Aas and Eikvil, 1999; Wang and Li, 2011; Yang and Pedersen, 1997). A growing number of machine learning techniques have been applied to text classification and some of them have proven to be successful (Miao and Kamel, 2011; Sebastiani, 2002). In the machine learn- ing approach, the learning process is an instance of supervised or semi-supervised learning because classifier can be built automatically by learning from adequately pre-labeled training documents and then classified unseen documents (Feldman and Sanger, 2007). However, the task of manual- ly labeling a large amount of documents is time- consuming and even impractical. Given a gen- eral classification task, people usually construc- t training data in two ways. One is augment- ing a small number of labeled documents with large amounts of unlabeled ones to guide the learn- ing model iteratively, so that the new classifier can label the unlabeled documents (Jiang, 2009; Nigam et al., 2000). In such studies, the boot- strapping technique is often used to label the un- labeled documents and refine the initial classifier (Gliozzo et al., 2009; Ko and Seo, 2009). The other is collecting training corpora from the We- b. Such works use the class name and its asso- ciated terms to collect training corpora iteratively (Huang et al., 2005). Cheng (2009) and Day et al. (2009) firstly sample the Web with a …

Content truncated to 500 words