Review on the Usage of Deep Learning Models in Multi-modal Sentiment Analysis

Abstract

In recent years, Sentiment Analysis is reshaping the business operations of many organizations by monitoring their brand reputation on social media and acquiring insights from customer"s feedback. Sentiment Analysis is one of the classification tools that identifies and extracts the subjective information of a product. This subjective information can be stated in different ways, such as feedback, discussions, blogs, podcasts, and video logs. This type of information generated by the empowered customers is known as user-generated content, which is traditionally in the form of words. The analysis was performed on a huge number of words using Natural Language Processing (NLP), which is a Unimodal Sentiment Analysis. With the rapid growth in the usage of the Internet, social media turned out to be a platform to share the thoughts of the individuals. This caused researchers to migrate from the traditional Unimodal analysis to Multimodal Sentiment Analysis, which includes video, audio, and images. This approach leverages the use of emotion and content and helps identify the scope and polarity of an individual’s sentiment. With the latest deep learning algorithms, Multimodal Sentiment Analysis can solve the problem of sarcasm identification. Multi-Modal Sentiment Analysis generates more accurate results compared to Uni Modal Sentiment Analysis. Therefore, this study aimed to define Sentiment Analysis and review the approaches and techniques in Sentiment Analysis from conventional Unimodal to Multimodal. In addition, this paper discusses a Multimodal Sentiment Analysis architecture using a transformers attention net.