E-commerce reviews are often huge in number, scattered in opinion, and full of repetition. To handle this, I design a layered framework that first clusters reviews with Aspect-Based Sentiment Analysis (ABSA) and then generates summaries in two stages. A weak-reference evaluation is added by rewriting official references from dataset AmaSum with an Large Language Model, so score bias from structural mismatch is reduced.

In this project, I'll put into practice: Aspect clustering that maintains opinions while filtering out noise Two-stage generating pipelines with carefully compared quality and runtime trade-offs Semantic and reference metrics (BERTScore and G-Eval) are inadequate for evaluating abstractive outputs fairly. Juppyter notebook that can summarise 500 real-world products on a single GPU