A Review of Transfer-Based Sentiment Classification

Abstract

Sentiment Analysis (SA) or opinion mining has become an important task in Natural Language Processing (NLP) to extract subjective information in a text. The conventional machine learning strategies demand a huge labelled dataset and require the same distribution of data across the training and testing data. Nevertheless, in practice, real world scenarios define domain transitions, low resource languages, multimodal inputs and dynamically evolving data environments. Transfer Learning (TL) is used to solve these issues by transfer knowledge in source domains to target domains. This review paper presents a comprehensive survey of sentiment analysis approaches based on transfer learning and CNN. It synthesizes insights from recent studies on deep transfer learning, multilingual transformers, multimodal sentiment analysis, aspect-based sentiment analysis (ABSA), and low-resource language adaptation. The paper typifies transfer learning techniques, describes important architecture, i.e., BERT and multilingual transformers, evaluates the performance gains, identifies the issues of negative transfer, and suggests the research directions. Due to the fast advancement of the Internet sector, sentiment analysis has emerged as one of the trending fields in natural language processing. Through it, the implicit emotion in the text can be effectively mined, which can help enterprises or organizations to make an effective decision, and the explosive growth of data undoubtedly brings more opportunities and challenges to the sentiment analysis. Transfer learning seeks to enhance the performance of target learners in specific domains by utilizing the knowledge derived from various but related source domains. This approach minimizes the reliance on a substantial amount of target domain data for building target learners