End-to-End Aspect-Guided Review Summarization at Scale

TL;DR

The paper tackles the challenge of surfacing concise, sentiment-grounded product feedback from large, noisy review streams. It introduces an end-to-end aspect-guided summarization pipeline that combines ABSA with guided prompts to produce summaries anchored in actual customer feedback, while being model-agnostic and scalable. Key contributions include a production-ready architecture, evidence from a large-scale online A/B test showing engagement improvements, and an open dataset of 11.8 million reviews with extracted aspects and summaries. The work demonstrates that grounding summaries in frequent aspects and representative reviews can improve user comprehension and engagement on e-commerce platforms.

Abstract

We present a scalable large language model (LLM)-based system that combines aspect-based sentiment analysis (ABSA) with guided summarization to generate concise and interpretable product review summaries for the Wayfair platform. Our approach first extracts and consolidates aspect-sentiment pairs from individual reviews, selects the most frequent aspects for each product, and samples representative reviews accordingly. These are used to construct structured prompts that guide the LLM to produce summaries grounded in actual customer feedback. We demonstrate the real-world effectiveness of our system through a large-scale online A/B test. Furthermore, we describe our real-time deployment strategy and release a dataset of 11.8 million anonymized customer reviews covering 92,000 products, including extracted aspects and generated summaries, to support future research in aspect-guided review summarization.

Figures (6)

• Figure 1: Product Summary example from wayfair.com
• Figure 2: The figure shows an example output generated by the system. It illustrates the aspect-guided review summarization pipeline, which extracts and consolidates aspect–sentiment pairs from customer reviews to identify the top 5 most frequent aspects. The system then generates a product summary using prompts built from representative reviews for each selected aspect-sentiment pair. The output includes selected aspects with counts and supporting reviews, along with a summary grounded in them.
• Figure 3: Illustration of the aspect extraction and consolidation process. Starting from a raw customer review, fine-grained aspect-sentiment pairs are extracted. These aspects are then mapped to broader canonical terms, resulting in a consolidated set of aspect-sentiment pairs.
• Figure 4: Prompt template for aspect extraction.
• Figure 5: Prompt template for aspect consolidation.