LLM-infused multi-module transformer for emotion-aware sentiment analysis in few-shot scenarios

Sentiment analysis, particularly in few-shot scenarios and under constraints of limited data availability, presents significant challenges in accurately capturing the nuanced emotions conveyed in online reviews and public opinions. To address these limitations, this study introduces the Cognemotive Transformer (CogTrans), an advanced model that integrates emotion-cognitive reasoning with transformer-based generative approaches to enhance sentiment analysis. The proposed CogTrans framework consists of four key modules. The Quantity Augmentation Module utilizes large language models (LLMs) to generate synthetic data, thereby improving learning efficiency in few-shot settings. The Emotional Cognitive Analysis (ECA) Module constructs a sentence–emotion tree to facilitate a deeper understanding of sentiment contexts. The Transformer-based Semantic Representation (T-SR) Module employs a mask-transformer architecture to extract high-quality semantic features. Lastly, the Crisis Entity and Intent Prediction (CEIP) Module leverages natural language processing (NLP) techniques to identify critical entities in crisis-related texts and infer their underlying intentions using COMET-ATOMIC 2020. The integration of these components significantly enhances sentiment prediction, particularly in noisy and data-scarce environments. Experimental evaluations demonstrate that CogTrans outperforms existing models in both sentiment classification and interpretability, achieving state-of-the-art results across multiple benchmark datasets. Its ability to provide well-contextualized sentiment predictions while incorporating emotional context, cognitive reasoning, and crisis-relevant insights makes it a highly promising tool for practical applications in crisis management and review analysis. • CogTrans integrates emotion-cognition with transformer-based sentiment models. • LLM-driven data augmentation improves few-shot sentiment classification. • Sentence-emotion trees enable deeper contextual emotional understanding. • Mask-transformer module extracts high-quality semantic representations. • Crisis-aware intent prediction supports analysis of urgent public discourse.