Cross-Domain Sentiment Analysis: A Multi-Task Learning Approach with Shared Representations

Abstract

This research aims to evaluate the use of Multi-Task Learning (MTL) in sentiment analysis in various domains. The primary limitation of dominant models of sentiment classification lies in the fact that most of them are domain-specific and because of assorted styles of language usage, context sensitivity and users’ behavior they fail to work across different domains. As pointed out earlier, transferring knowledge and feature learning are the key steps in transferring learnt models from one domain to the other in handling these challenges, this study aims at adopting an MTL based approach for better representation of a shared representation that can foster better generalization of the model across the different domains. This way, the proposed model allows for establishing the mixture of separating and common features for sentiment analysis across the contextual domains, with minor tuning of the model on each of them. These experiments prove that the accuracy, precision, recall, and F1 scores of MTL-shared are superior to those of the traditional single-domain and domain-adaptation models. The model also has high performance in cross-domain as seen in evaluation on unseen domains, and hence it can be a useful method in real-world circumstances where sentiment analysis must be conducted in diverse domains.