SHE: Sentiment Hashtag Embedding Through Multitask Learning

Abstract

Recent studies have shown the importance of utilizing hashtags for sentiment analysis task on social media data. However, as the hashtag generation process is less restrictive, it throws several challenges, such as hashtag normalization, topic modeling, and semantic similarity. Recently, researchers have tried to address the above-mentioned challenges through representation learning. However, most of the studies on hashtag embedding try to capture the semantic distribution of hashtags and often fail to capture the sentiment polarity. Furthermore, generating a task-specific hashtag embedding can distort its semantic representation, which is undesirable for sentiment representation of hashtag. Therefore, this article proposes a semisupervised sentiment hashtag embedding (SHE) model, which is capable of preserving both semantic as well as sentiment distribution of the hashtags. In particular, SHE leverages a multitask learning approach using an autoencoder and a convolutional neural network-based classifier. To assess the efficacy of hashtag embedding, we compare the performance of SHE against suitable baselines for three different tasks, namely, hashtag sentiment classification, tweet sentiment classification, and retrieval of semantically similar hashtags. It is evident from various experimental results that SHE outperforms the majority of the baselines with significant margins.