A topic BiLSTM model for sentiment classification

Abstract

The Long Short Term Memory (LSTM) network is very effective for capturing sequence information which can help to analyze sentiments. However, it fails to capture the meaning of polysemous word under different contexts. In this paper, we propose topic information-based bidirectional LSTM (BiLSTM) model for sentiment classification. BiLSTM model learns topic information to obtain the sensitive representation of the polysemous word under given circumstance. The topic information is generated through a topic modeling via Latent Dirichlet Allocation (LDA). The topic information-based BiLSTM network allows the model to capture the meaning of the polysemous word and long sequence information automatically. The experimental results on real-world datasets demonstrate that the proposed method outperforms the task of benchmark sentiment classification on SemEval 2013 and IMDB.