Cross Language Text Classification by Model Translation and Semi-Supervised Learning

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 1057–1067, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Cross Language Text Classification by Model Translation and Semi-Supervised Learning Lei Shi Yahoo! Global R&D Beijing, China lshi@yahoo-inc.com Rada Mihalcea University of North Texas Denton, TX, U.S.A. rada@cs.unt.edu Mingjun Tian Yahoo! Global R&D Beijing, China mingjun@yahoo-inc.com Abstract In this paper, we introduce a method that au- tomatically builds text classifiers in a new lan- guage by training on already labeled data in another language. Our method transfers the classification knowledge across languages by translating the model features and by using an Expectation Maximization (EM) algorithm that naturally takes into account the ambigu- ity associated with the translation of a word. We further exploit the readily available un- labeled data in the target language via semi- supervised learning, and adapt the translated model to better fit the data distribution of the target language. 1 Introduction Given the accelerated growth of the number of mul- tilingual documents on the Web and elsewhere, the need for effective multilingual and cross-lingual text processing techniques is becoming increasingly im- portant. There is a growing number of methods that use data available in one language to build text pro- cessing tools for another language, for diverse tasks such as word sense disambiguation (Ng et al., 2003), syntactic parsing (Hwa et al., 2005), information re- trieval (Monz and Dorr, 2005), subjectivity analysis (Mihalcea et al., 2007), and others. In this paper, we address the task of cross-lingual text classification (CLTC), which builds text classi- fiers for multiple languages by using training data in one language, thereby avoiding the costly and time- consuming process of labeling training data for each individual language. The main idea underlying our approach to CLTC is that although content can be expressed in different forms in different languages, there is a significant amount of knowledge that is shared for similar topics that can be effectively used to port topic classifiers across languages. Previous methods for CLTC relied mainly on ma- chine translation, by translating the training data into the language of the test data or vice versa, so that both training and test data belong to the same lan- guage. Monolingual text classification algorithms can then be applied on these translated data. Al- though intuitive, these methods suffer from two ma- jor drawbacks. First, most off-the-shelf machine translation sys- tems typically generate only their best translation for a given text. Since machine translation is known to be a notoriously hard problem, applying mono- lingual text classification algorithms directly on the erroneous translation of training or test data may severely deteriorate the classification accuracy. Second, similar to domain adaptation in statisti- cal machine learning, due to the discrepancy of data distribution between the training domain and test do- main, data distribution across languages may vary because of the difference of culture, people’s inter- ests, linguistic expression in different language re- gions. So even if the translation of training or …

Content truncated to 500 words