Joint Embedding of Emoticons and Labels Based on CNN for Microblog Sentiment Analysis
Yongcai Tao, Xinqian Zhang, Lei Shi, Lin Wei, Zhaoyang Hai, Junaid Abdul Wahid · 2019
Microblog sentiment analysis is a technique to analyze the emotion types of microblog texts, which is beneficial to control the direction of the public's attitudes or emotions. Different researches are being developed to guide sentiment analysis. However, these methods do not consider or do not give enough attention to emoticons and labels present in the microblog sentences, which cause a loss of potential information. To solve this problem, the research proposes a novel model, called joint embedding of Emoticons and Labels based on CNN (EL-CNN). EL-CNN applies an input encoder to convert input sequences into fix-length semantic vectors, a label encoder to convert labels into fix-length semantic vectors and a matcher to calculate a matching score between the fix-length semantic vectors obtained from the input encoder and the label encoder, which turns the sentiment classification tasks into vector matching tasks. The work maintains the interpretability of word embedding and enjoys a built-in ability to capture sentiment information from emoticons and labels. Experimental results on two public microblog benchmark corpora (NLPCC2013 and NLPCC2014) show that the EL-CNN model achieves significant improvements as compared to all start-of-the-art methods.