Encoding Category Correlations into Bilingual Topic Modeling for Cross-Lingual Taxonomy Alignment

Abstract

Cross-lingual taxonomy alignment (CLTA) refers to mapping each category in the source taxonomy of one language onto a ranked list of most relevant categories in the target taxonomy of another language. Recently, vector similarities depending on bilingual topic models have achieved the state-of-the-art performance on CLTA. However, these models only model the textual context of categories, but ignore explicit category correlations, such as correlations between the categories and their co-occurring words in text or correlations among the categories of ancestor-descendant relationships in a taxonomy. In this paper, we propose a unified solution to encode category correlations into bilingual topic modeling for CLTA, which brings two novel category correlation based bilingual topic models, called CC-BiLDA and CC-BiBTM. Experiments on two real-world datasets show our proposed models significantly outperform the state-of-the-art baselines on CLTA (at least +10.9% in each evaluation metric).