Cross-Lingual Sentiment Classification via Bi-view Non-negative Matrix Tri-Factorization

Recently the sentiment classification problem interests the researchers over the world, but most sentiment corpora are in English, which limits the research progress on sentiment classification in other languages. Cross-lingual sentiment classification aims to use annotated sentiment corpora in one language (e. g. English) as training data, to predict the sentiment polarity of the data in another language (e. g. Chinese). In this paper, we design a bi-view non-negative matrix tri-factorization (BNMTF) model for the cross-lingual sentiment classification problem. We employ machine translation service so that both training and test data is able to have two representation, one in source language and the other in target language. Our BNMTF model is derived from the non-negative matrix tri-factorization models in both languages in order to make more accurate prediction. Our BNMTF model has three main advantages: (1) combining the information from two views (2) incorporating the lexical knowledge and training document label knowledge (3) adding information from test documents. Experimental results show the effectiveness of our BNMTF model, which can outperform other baseline approaches to cross-lingual sentiment classification.