Chinese Short Text Classification Based on Domain Knowledge

International Joint Conference on Natural Language Processing, pages 859–863, Nagoya, Japan, 14-18 October 2013. Chinese Short Text Classification Based on Domain Knowledge Xiao Feng Institute of Automation Chinese Academy of Science xiao.feng@ia.ac.cn Yang Shen State Administration for In- dustry & Commerce of the People's Republic of China shenyang@saic.gov.cn Chengyong Liu Information Center of Gen- eral Administration of Press and Publication of PR China liucy_gapp@sina.com Wei Liang Institute of Automation Chinese Academy of Science wei.liang@ia.ac.cn Shuwu Zhang Institute of Automation Chinese Academy of Science shuwu.zhang@ia.ac.cn Abstract People are generating more and more short texts. There is an urgent demand to classify short texts into different domains. Due to the shortness and sparseness of short texts, con- ventional methods based on Vector Space Model (VSM) have limitations. To tackle the data scarcity problem, we propose a new mod- el to directly measure the correlation between a short text instance and a domain instead of representing short texts as vectors of weights. We firstly draw domain knowledge for each user-defined domain using an external corpus of longer documents. Secondly, the correlation is calculated by measuring the proportion of the overlapping part of the instance and the domain knowledge. Finally, if the correlation is greater than a threshold, the instance will be classified into the domain. Experimental re- sults show that the classifier based on the pro- posed model outperforms the state-of-the-art baselines based on VSM. 1 Introduction In recent years, web services are generating more and more short texts including micro-blogs, cus- tomer reviews, chat messages and so on. Howev- er, a user is often only interested in very small part of these data. There is an urgent demand to classify incoming short texts into different do- mains, so that users are not overwhelmed by the raw data. As short texts do not provide sufficient word occurrences (i.e., the length of a micro-blog is limited to 140 characters), conventional text classifiers often cannot achieve high accuracy, especially when the number of training examples is small. Vector Space Model (VSM) is a very popular document representation model, where each doc- ument is represented as a vector of weights. Text classification methods based on VSM perform well when processing documents in regular length (Berry and Michael, 2004). But, the sparsity of VSM will reduce the classification accuracy when processing short texts. There have been several studies that attempted to solve the problem of data sparseness in VSM. One way is to select more useful features using additional semantics from Wikipedia (Banerjee et al. ,2007), WordNet (Hu et al , 2009) or HowNet (Liu et al. , 2010). Another way is to expand the coverage of classifier by using back- ground knowledge drawn from much longer ex- ternal data sources. Zelikovitz and Hirsh (2000) utilized a corpus of unlabeled longer documents as a “bridge”, to connect the test example with training examples. Phan et al. (2008) and Chen et al. (2011) integrated the original short text with hidden topics discovered from external large- scale data collections …

Content truncated to 500 words