Sentiment Strength Prediction Using Auxiliary Features

With an increasingly large amount of sentimental information embedded in online documents, sentiment analysis is quite valuable to product recommendation, opinion summarization, and so forth. Different from most works on identifying documents' qualitative affective information, this research focuses on the measurement of users' intensity over each sentimental category. Affect indicates positive or negative sentiment, while cognition includes certainty and tentative. Thus, our research can help bridge the cognitive and affective gaps between users and documents. The contributions of this study are twofold: (i) we proposed a neural network-based framework to sentiment strength prediction by convolving hybrid vectors, and (ii) we considered words jointly with a set of linguistic features for enhancing model robustness and adaptiveness. By exploiting the auxiliary features of sentiments from the corpus, the proposed model did not rely on well-established lexicons, and showed its robustness over sparse words. Experiments on six corpora validated the effectiveness of our sentiment strength prediction method.