Recent Advancements and Challenges in Multimodal Sentiment Analysis: A Survey
Zhen-Xun Dong, Han Liu · 2023
Sentiment analysis aims to extract people's underlying attitudes, opinions and thoughts toward various topics, products and services. Sentiment analysis is a popularly adopted technique that involves applications across domains such as business, healthcare, government and marketing. Traditional sentiment analysis methods mainly focus on text contents. However, recent technological advancements have enabled the expression of opinions through different forms of media, such as images or audios, leading to sentiment analysis using multiple modalities rather than text alone. Multimodal sentiment analysis results in a more comprehensive understanding of the emotions and opinions expressed in a given content by combining information from different modalities. A model of multimodal sentiment analysis involves three main parts, namely, unimodal feature extraction, multimodal fusion and sentiment polarity prediction. This paper provides a comprehensive overview of multimodal fusion methods along with their corresponding model architectures. Moreover, several more recently released models are introduced and the strengths and limitations of these models are discussed comparatively. Finally, some challenges and future research directions are identified.