Deep Bi-directional Long Short-Term Memory Neural Networks for Sentiment Analysis of Social Data

Abstract

Sentiment analysis (SA) has been attracting a lot of studies in the field of natural language processing and text mining. Recently, there are many algorithm’s enhancements in various SA applications are investigated and introduced. Deep Convolutional Neural Networks (DCNNs) have recently been shown to give the state-of-the-art performance on sentiment classification of social data. Although, these solutions effectively address issues of multi-levels features presentation but having some limitations of temporal modeling. In addition, the Bi-directional Long Short-Term Memory (BLTSM) conventional models have encountered some limitations in presentation with multi-level features but can keep track of the temporal information while enabling deep representations in the data. In this paper, we propose to use Deep Bi-directional Long Short-Term Memory (DBLSTM) architecture with multi-levels feature presentation for sentiment polarity classification (SPC) on social data. By using DBLSTM, we can exploit more level features than BLTSM and inherit temporal modeling in BLTSM. Moreover, the language of social data is very informal with misspellings and abbreviations. One word can be appeared in multiple formalities, which is a challenge in word-level models. We use character-level as input of DBLSTM neural network (called Character DBLSTM - CDBLSTM) for learning sentence level presentation. The experimental results show that the performance of our model is competitive with state-of-the-art of SPC on Twitter’s data. Our model achieves 85.86 % accuracy on Stanford Twitter Sentiment corpus (STS) and 84.82 % accuracy on the subtasks B of SemEval-2016 Task 4 corpus.