Emotion quantification and classification using the neutrosophic approach to deep learning

Abstract

Advancements in the rapidly evolving specialization of deep learning have aided in improving several natural language understanding tasks. Sentiment and emotion classification models have improved, but when it comes to fine-grained sentiment analysis, these models can perform better. Human sentiment in natural language is generally an intricate combination of emotions, which can sometimes be indeterminate, neutral, or ambiguous. In the case of fine-grained sentiment analysis, the sentiments can be very similar to each other and interconnected, e.g., anger and fear. Most deep learning systems try to solve the problem of fine-grained sentiment analysis as a classification problem. However, fine-grained sentiments might combine similar emotions with one primary emotion. Trying to solve the problem as a classification task can result in better performance on benchmarks but does not ensure a better understanding and representation of language. The proposed work explores applying neutrosophy for fine-grained sentiment analysis using large language models. Neutrosophy identifies neutralities and employs membership functions (neutral, positive, negative) to quantify an instance into Single Valued Neutrosophic Sets (SVNS). This paper introduces Refined Emotion Neutrosophic Sets (RENS) for emotions (with four emotions) and Refined Ekman’s Emotion Neutrosophic Sets (REENS) with seven emotions. In this paper, refined neutrosophic sets with membership functions are employed for each sentiment across a given taxonomy and assigned their values using the Neutrosophic Iterative Neural Clustering (NINC) algorithm proposed in this paper. It facilitates not only classifying sentiments but also quantifying the presence of each sentiment present in a given sample. It aids in better understanding and representation of samples across multiple sentiments, as in fine-grained sentiment analysis, experiments are performed on the GoEmotions dataset. The proposed approach performs on par with cross-entropy deep learning classifiers and is reproducible across different pre-trained language models.