Breaking Barriers in Sentiment Analysis and Text Emotion Detection: Toward a Unified Assessment Framework

Sentiment analysis (SA) and text emotion detection (TED) are two computer techniques used to analyze text. SA categorizes text into positive, negative, or neutral opinions, while TED can identify a wide array of emotional states, allowing an automated agent to respond appropriately. These techniques can be helpful in areas such as employee and customer management, online support, and customer loyalty, where identifying human emotions is crucial. Among other approaches, research has been conducted using machine learning (ML) algorithms, and labeled datasets have been created to train these models. Current state-of-the-art research for supervised ML algorithms reports good performance for TED (approximately 80% accuracy) and even better results for SA (above 90%). After conducting an extensive review of 30 surveys, the primary objective of this manuscript is to point out that most of these articles (94%) focus heavily on comparing the applied computational methods (the algorithm). At the same time, relatively diminished attention is paid to three other critical factors, namely the selection of an appropriate emotion model (mentioned only in 23% of cases), the corpora utilized for training (30%), and the data source employed during analysis and evaluation (20%). The lack of standardization across these essential elements presents a significant challenge when performing meaningful performance comparisons among algorithms. Consequently, the absence of a unified framework for comparison hampers the practical implementation of SA and TED techniques within mission-critical scenarios within real-world mission-critical scenarios.

Automated summary

Sentiment analysis and text emotion detection are computer techniques used to analyze text and identify human emotions. However, most research focuses on comparing computational methods, while less attention is paid to selecting emotion models, training corpora, and data sources. The lack of standardization in these critical factors presents a challenge for meaningful performance comparisons among algorithms.