Large language model-based customer feedback analysis in hospitality domain for improvement suggestions

Abstract

This research investigates the challenge of systematically transforming customer feedback into actionable service improvement suggestions in the hospitality sector, where unstructured reviews often go underutilized. To address this, we propose a novel threestage natural language processing (NLP) pipeline that utilizes fine-tuned large language models (LLMs) designed for efficiency, scalability, and domain adaptation. The study uses a dataset of approximately 4,300 recent reviews collected from TripAdvisor and Booking.com to train and evaluate each stage of the pipeline. First, a sentiment classification model (DistilBERT-base-uncased) categorizes reviews as fully positive or negative. Next, a fine-tuned sequence-to-sequence model (Flan-T5-base) extracts key negative aspects. Finally, an open-ended suggestion generation stage is performed using a fine-tuned LLM (Mistral-7B-Instruct). Unlike prior approaches that focus narrowly on sentiment or rely on manual rule engineering, our pipeline offers a fully automated, end-to-end framework tailored for high-volume review environments. The system is evaluated using standard metrics. The sentiment classification stage achieved over 92% accuracy with high precision, recall, and F1-scores. The aspect extraction stage attained a BLEU score of 0.65 and ROUGE-L of 0.85. The final suggestion generation stage produced coherent outputs. Although the BLEU score was low due to mismatch with references, human evaluation confirmed the suggestions were contextually appropriate. This study contributes to applied NLP by demonstrating how lightweight, fine-tuned models can be integrated to produce coherent, domain-specific, and scalable insights for real-world service improvement. The approach opens avenues for further research in multi-domain adaptation, real-time feedback analytics, and explainable AI for recommendation systems.