Incorporating Positional Information into Deep Belief Networks for Sentiment Classification

Deep belief networks (DBNs) have proved powerful in many domains including natural language processing (NLP). Sentiment classification has received much attention in both engineering and academic fields. In addition to the traditional bag-of-word representation for each sentence, the word positional information is considered in the input. We propose a new word positional contribution form and a novel word-to-segment matrix representation to incorporate the positional information into DBNs for sentiment classification. Then, we evaluate the performance via the total accuracy. Consequently, our experimental results show that incorporating positional information performs better on ten short text data sets, and also the matrix representation is more effective than the linear positional contribution form, which further proves the positional information should be taken into account for sentiment analysis or other NLP tasks.