Toward multi-label sentiment analysis: a transfer learning based approach

• Published in: Journal of big data 2020-01, Vol.7 (1), p.1-26, Article 1
• Main Authors: Tao, Jie, Fang, Xing
• Format: Article
• Language: English
• Subjects: Big Data; Classification; Communications Engineering; Computational Science and Engineering; Computer Science; Contrarian investing; Data mining; Data Mining and Knowledge Discovery; Database Management; Deep learning; Digital media; Domains; Empirical analysis; Information Storage and Retrieval; Learning; Mathematical Applications in Computer Science; Multi-label classification; Natural language processing; Networks; Sentiment analysis; Transfer learning
• ISSN: 2196-1115; 2196-1115
• DOI: 10.1186/s40537-019-0278-0

Sentiment analysis is recognized as one of the most important sub-areas in Natural Language Processing (NLP) research, where understanding implicit or explicit sentiments expressed in social media contents is valuable to customers, business owners, and other stakeholders. Researchers have recognized that the generic sentiments extracted from the textual contents are inadequate, thus, Aspect Based Sentiment Analysis (ABSA) was coined to capture aspect sentiments expressed toward specific review aspects . Existing ABSA methods not only treat the analytical problem as single-label classification that requires a fairly large amount of labelled data for model training purposes, but also underestimate the entity aspects that are independent of certain sentiments. In this study, we propose a transfer learning based approach tackling the aforementioned shortcomings of existing ABSA methods. Firstly, the proposed approach extends the ABSA methods with multi-label classification capabilities. Secondly, we propose an advanced sentiment analysis method, namely Aspect Enhanced Sentiment Analysis (AESA) to classify text into sentiment classes with consideration of the entity aspects. Thirdly, we extend two state-of-the-art transfer learning models as the analytical vehicles of multi-label ABSA and AESA tasks. We design an experiment that includes data from different domains to extensively evaluate the proposed approach. The empirical results undoubtedly exhibit that the proposed approach outperform all the baseline approaches.