Sentence-Level Sentiment Analysis Using a CNN Model on Contextualized Word Representations

Abstract

The progress in Internet-based applications such as websites and social networks has resulted in the generation of significant amounts of reviews and opinions associated with policies, influencers, services, and products. The emotion in these opinions becomes a significant source for politicians, celebrities, servicers, and producers who intend to improve their activities and make better decisions. Sentiment analysis (SA) is an essential tool for extracting and analyzing these emotions. Various state-of-the-art SA methods with good performance have been developed. This paper presents a model, called context-based convolutional neural network (CNN), for sentence-level SA. Unlike previous methods, the proposed model focuses on the role of context information expressed in a sentence for SA. The proposed method comprises the following procedures: (i) Sentences are converted into vectors based on the BERT model; (ii) contextualized word representations are extracted using the BiLSTM model over word embeddings; (iii) sentence-level sentiment is analyzed using the CNN model over the contextualized word representations. Experiments on benchmark datasets prove that the proposed method can improve the performances of previous methods.