Critical Analysis of Limitations in Contemporary Sentiment Analysis Techniques: A Foundation for Multimodal Deep Learning Approaches

Sentiment analysis has grown in importance in the field of natural language processing (NLP) due to its use in marketing, political predictions, public opinion research, and many other fields. However, the most recent techniques employed for sentiment analysis are still unsatisfactory. The limitations arise in sarcasm detection, classification of the context as culturally nuanced or domain dependent, reliance on text data. In this paper we study different approaches used for sentiment analysis including lexicon-based methods and deep learning techniques focusing first on their strengths and weaknesses when applied to real-world datasets and benchmark studies. The results reveal that attempts at capturing complex emotional expression failure strongly suggesting simplistic approach relies too heavily on a single modality.