Sentiment analysis and summarization with ChatGPT: implications for sales prediction

Abstract Recent advancements in Large Language Models (LLMs) have enabled their adoption across a wide range of business applicationshave facilitated their widespread adoption in diverse business applications. ChatGPT, in particular, exemplifies these advancements with its exceptional capabilities in interpreting contextual and complex information and generating natural language response. Using Booking.com review datasets and government tax revenue data, this study compares ChatGPT's predictive performance with traditional methods (review ratings and machine learning approaches) for sales revenue sentiment analysis. Moreover, this study also examines ChatGPT's summarization efficacy in extracting sales-predictive information while reducing redundancy and noise. The objective is to address the gap in the application of LLMs in information processing and summarization in emotional and heuristic contexts. This addresses LLM application gaps in emotional/heuristic contexts for information processing and summarization. We first benchmark review ratings, RoBERTa, and ChatGPT for sentiment index extraction in sales prediction. Furthermore, we reconstruct a ChatGPT-summarized review dataset, introducing novel prompt engineering methods for sentiment analysis and user-content summarization. We additionally examine the effectiveness of ChatGPT-based summaries by highlighting the difference between raw and summarized reviews, fine-tune the RoBERTa model with review ratings, and employ a few-shot approach to fine-tune ChatGPT, aiming to evaluate the importance of reduced redundancy and information noise in using user-generated content for sales prediction.