Regularization of multilingual topic models

Abstract

A multilingual probabilistic topic model based on the additive regularization ARTM allowing to combine both a parallel or comparable corpus and a bilingual translation dictionary is proposed. Two approaches to include information from a bilingual dictionary are discussed: the first one takes into account only the fact of connection between word translations, whereas the second one learns the translation probabilities for each topic. To measure the quality of the proposed multilingual topic model, a cross-language search is performed. For each query document in one language, it is found its translation on an other language. It is shown that the combined translation of words from a bilingual dictionary and the corresponding connected documents improves the cross-lingual search compared to the models using only one information source. The use of learning word translation probabilities for bilingual dictionaries improves the quality of the model and allows one to determine a context (a set of topics) for each pair of word translations, where these translations are appropriate.