Inverted Bilingual Topic Models for Lexicon Extraction from Non-parallel Data

Abstract

A good lexicon is an important resource for various cross-lingual tasks such as information retrieval and text mining. In this paper, we focus on extracting translation pairs from non-parallel cross-lingual corpora. Previous lexicon extraction algorithms for non-parallel data generally rely on an accurate seed dictionary and extract translation pairs by using context similarity. However, there are two problems. One, a lot of semantic information is lost if we just use seed dictionary words to construct context vectors and obtain the context similarity. Two, in practice, we may not have a clean seed dictionary. For example, if we use a generic dictionary as a seed dictionary in a special domain, it might be very noisy. To solve these two problems, we propose two new bilingual topic models to better capture the semantic information of each word while discriminating the multiple translations in a noisy seed dictionary. We then use an effective measure to evaluate the similarity of words in different languages and select the optimal translation pairs. Results of experiments using real Japanese-English data demonstrate the effectiveness of our models.