Chinese-thai cross-language topic extraction and alignment

Cross-language topic extraction aims at extracting related topics from texts in different languages. A widely used approach is to build a topic model from a set of parallel or comparable texts in different languages. However, for resource-lean languages such as Thai, such texts may be unavailable. An alternative approach is to construct separate topic models for different languages, and then to align them. We investigate this approach in this paper for Chinese-Thai texts. A key problem to align topics in different languages is to estimate their similarity, which requires the translation of topic words between the two languages. Unfortunately, the translation between Chinese and Thai is not accurate enough for this task. Therefore, we use English as pivot language, through which the topics in Chinese and Thai are aligned. In addition, the representative topic words are filtered so that only those that are specific for the topic are kept. Experimental results performed on a set of news articles in Chinese and Thai show that our approach achieves a higher effectiveness than a traditional method.