Unsupervised Bilingual Sentiment Word Embeddings for Cross-lingual Sentiment Classification

Abstract

In recent years, bilingual word embeddings have been used to promote sentiment classification task in low-resource languages. However, existing bilingual word embedding methods either require annotated cross-lingual data or fail to capture enough sentiment information. In this paper, we propose Unsupervised Bilingual Sentiment word Embeddings (UBSE), which only need source-language annotated corpora and a monolingual sentimental lexicon. This method is constructed in an unsupervised way, we pre-train a projection matrix between the source and target languages using Generative Adversarial Nets (GAN) without using any parallel corpora. Further, we incorporate a monolingual sentiment lexicon from the source language to fine-tune the model, making it more sensitive to sentiment implication. Experiments on Spanish, Catalan and Basque demonstrate that the proposed approach on sentence-level cross-lingual sentiment classification significantly outperforms competitive baseline models which use cross-lingual dictionaries, even comparable with translation based methods.