Analyzing digital consumer insights through RoBERTa LLM based sentiment analysis and topic modeling.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Consumer Behavior Analysis

Background:

• Customer reviews are vital for understanding consumer sentiment and purchasing trends in online shopping.
• Traditional sentiment analysis methods struggle with the complexity and context of textual data.
• Large language models offer advanced capabilities for nuanced text analysis.

Purpose of the Study:

• To improve sentiment classification performance in online consumer reviews.
• To leverage RoBERTa-Large's contextual embeddings and attention mechanisms for enhanced analysis.
• To identify key themes in consumer feedback using topic modeling.

Main Methods:

• Applied RoBERTa-Large for sentiment classification, utilizing its contextual embeddings and attention mechanisms.
• Employed Latent Dirichlet Allocation (LDA) for topic modeling on consumer feedback datasets.
• Conducted comparative analysis against traditional machine learning (TF-IDF) and deep learning models.
• Utilized SHAP and LIME for model interpretability and explanation.

Main Results:

• RoBERTa-Large achieved a superior accuracy of 93.59% in sentiment classification.
• The model significantly outperformed baseline methods, including traditional and other deep learning approaches.
• Topic modeling identified prevalent themes within consumer feedback data.
• Interpretability techniques provided transparent explanations for model predictions.

Conclusions:

• RoBERTa-Large significantly advances sentiment analysis accuracy for online consumer reviews.
• Contextual embeddings and attention mechanisms are crucial for capturing semantic nuances in text.
• The study highlights the potential of advanced NLP models for actionable business intelligence.
• Model interpretability methods enhance the trustworthiness and practical application of sentiment analysis findings.