Cross-lingual Text Classification Using Topic-Dependent Word Probabilities

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1466–1471, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Cross-lingual Text Classification Using Topic-Dependent Word Probabilities Daniel Andrade Akihiro Tamura Masaaki Tsuchida Kunihiko Sadamasa Knowledge Discovery Research Laboratories, NEC Corporation, Japan {s-andrade@cj, a-tamura@ah, m-tsuchida@cq, k-sadamasa@az}.jp.nec.com Abstract Cross-lingual text classification is a major challenge in natural language processing, since often training data is available in only one language (target language), but not avail- able for the language of the document we want to classify (source language). Here, we pro- pose a method that only requires a bilingual dictionary to bridge the language gap. Our proposed probabilistic model allows us to es- timate translation probabilities that are condi- tioned on the whole source document. The assumption of our probabilistic model is that each document can be characterized by a dis- tribution over topics that help to solve the translation ambiguity of single words. Us- ing the derived translation probabilities, we then calculate the expected word frequency of each word type in the target language. Fi- nally, these expected word frequencies can be used to classify the source text with any classi- fier that was trained using only target language documents. Our experiments confirm the use- fulness of our proposed method. 1 Introduction Text classification is ubiquitous in natural language processing. It’s applications range from simple topic detection, like articles about sport vs articles about computers, to sentimental analysis, and subtle dis- crimination of Tweets that report the abuse of drugs or the metaphoric use of drugs (“love is like a drug”). Text classification hugely relies on manually anno- tated training data in one language. However, creating training data for each language is expensive, and therefore, we are interested in us- ing training data given in only one language (e.g. English, denoted as target language) to classify text written in a different language (e.g. Japanese, denoted as source language).Chinese, or Our approach addresses this issue by using a sim- ple bilingual dictionary. Bilingual dictionaries have the great advantage that they are available often for free1, and have good coverage for major languages, like Chinese and Japanese. With the help of the dic- tionary, we calculate the expected frequency of each word in the target language. Finally, we create a fea- ture vector in the target language that is used as input for the text classifier. However, due to the translation ambiguity of a word in the source language, it is important to care- fully choose the translation probability for calculat- ing the expected frequencies of the target words. For example, consider a Japanese news article that con- tains the word 拘束 (restrict, restrain, in custody), and we want to find out whether the article is about “foreign policy” or not. The most simple method is to use all its English translations, and assume a uniform distribution over them, i.e. {0.33, 0.33 and 0.33}. However, depending on the topic of the news article, …

Content truncated to 500 words