Personalized Book Reviewer Simulation via LLMs

Bachelor & Master Thesis

Large Language Models (LLMs) have shown impressive capabilities in understanding and generating human-like text. One promising application is to simulate personalized recommendations and opinions, such as those found in book reviews. While platforms like Goodreads or Amazon provide abundant review data, they do not offer deeply personalized reviewer profiles that reflect consistent individual preferences, tone, or judgment. A fine-tuned LLM that captures the nuanced reviewing style of a specific person or group could provide more relatable and trustworthy recommendations to users.

This project explores the idea of constructing a personalized book reviewer using fine-tuned LLMs. By collecting and analyzing publicly available or user-provided book reviews, the goal is to fine-tune a base LLM (e.g., LLaMA, DeepSeek, or GPT-based models) to reflect a specific reviewing persona. The resulting model should be able to generate new book reviews or critique summaries in the style of the original reviewer. Evaluation will involve both automatic metrics (e.g., perplexity, stylistic similarity) and human judgment to assess how closely the model mimics the intended persona. The work also opens avenues for further personalization, such as aligning generated content with user-specific genre preferences, sentiment tendencies, or historical ratings.

Required Knowledge:

Python programming, experience with LLMs and fine-tuning techniques, basic understanding of NLP and evaluation metrics, interest in literature and recommendation systems

What you will do will include:

• Collect and preprocess book review data from selected sources or users.
• Fine-tune a suitable pre-trained LLM on the curated review dataset.
• Design a framework to simulate personalized review generation.
• Evaluate the model outputs using both automatic and manual methods.
• Explore enhancements such as style transfer, preference modeling, or integration with book metadata for more accurate personalization.

For more details about this topic, please contact Dr. Shengcheng Yu (shengcheng.yu[at]tum.de).