A hybrid neural network approach for fine-grained emotion classification and computing

The traditional sentiment classification focuses more on the three polarities of sentiments, which are not fine-grained enough to fully characterize the overall evolution of sentiments and face the problem of sparse features. Aiming at these limitations, we propose a novel method to handle this problem: a hybrid neural network model for fine-grained emotion classification and computing. First, a sentiment dictionary is constructed by using the sentiment lexical ontology. Then, according to dependency parsing, a textual sentiment classifier is built with the aid of long short term memory network technologies; fine-grained netizen sentiment index is calculated. Finally, our approach was applied to practical business problem-exploring the interactions between the netizen sentiment index and the stock return in order to test its reliability. The experimental results show that compared with traditional methods, this approach improves the accuracy of sentiment classification, possess higher classification performance, reduces the number of iterations and saves computing resources. The empirical analysis demonstrate that hybrid method is rapid, effective and feasible, could be more suitable for fine-grained emotion computing.