Not Only the Contextual Semantic Information: A Deep Fusion Sentimental Analysis Model Towards Extremely Short Comments

Liping Hua, Qinhui Chen, Zelin Huang, Hui Zhao, Gang Zhao

Abstract: Extremely short comments (ESC) often contain rich information to convey users’ emotions towards content. However, conducting sentiment analysis on ESC is challenging due to the limited contextual semantic information and colloquial expressions. Traditional methods mainly focus on contextual text features. In this work, we propose a novel model, named Chinese Phonetic-Attentive Deep Fusion Network (CPADFN) that attentively fuse the Chinese phonetic alphabet features of the ESC, meta-information about the ESC along with the contextual text features. First, the multi-head self-attention mechanism is utilized to obtain the phonetic alphabet representation and the sentence representation separately. Also, a fully-connected layer is used on the embeddings of the meta-information about the ESC to obtain the meta-information representation. Then, the local activation unit is employed to attentively fuse these feature representations. Bi-LSTM is applied to address the sequence dependency across these fused features separately. Third, a fully-connected layer with softmax function is applied to predict emotional labels. We conduct experiments on a self-crawled ESC dataset DanmuCorpus, and two public Chinese short text datasets, MovieReview and WeiboCorpus. The experimental results demonstrate that CPADFN achieves better performances.