Exploring Deep Recurrent Convolution Neural Networks for Subjectivity Classification

Abstract

Subjectivity classification aims to discriminate whether the given text presents opinions or states facts, which is an important foundation and component of sentiment analysis task. The characteristics of microblog (e.g., short text, full noise, casual expression, and incomplete grammatical structure) bring new challenges to microblog text subjectivity classification task. The existing subjectivity classification approaches of microblog text mainly depend on the techniques of artificially constructing emotional features that are commonly based on the bag-of-words model, which will lead to the curse of dimensionality and data sparsity problem. Furthermore, microblog text usually contains the left-handed compliment or implicit representation. It is difficult to find and describe this implicit semantic representation by artificially constructing emotional features, which seriously affects the subjectivity classification accuracy. To address these problems, we introduce a deep recurrent convolution neural network model for subjectivity classification of microblog text with the multi-features combination. In our model, we apply the convolution neural networks to automatically learn classification features and use the recurrent neural networks to abstract and memorize the sequence semantic relations of microblog text. We also employ pre-trained word embedding to reduce the representation dimension of the text, which avoids dimension disasters and data sparsity issues. We perform experiments on six commonly used Twitter datasets. The experimental results demonstrate that our proposed model outperforms the state-of-the-art approaches on several datasets. Compared with the baseline methods, the accuracy and the average F1-measure of our proposed model for subjectivity classification are improved by a minimum of 1.21% and 2.82%, respectively.