Fine-Grained Aspect-Based Sentiment Analysis for Meituan's To‑Restaurant Business

Background

Meituan, a large online life‑service platform, aims to link consumers and merchants through technology. The "to‑restaurant" (到餐) business generates massive user‑generated content (UGC) in the form of reviews. Extracting fine‑grained sentiment (aspect, opinion, polarity) from these reviews can help users make decisions and assist merchants in monitoring service quality.

Problem Definition

Fine‑grained sentiment analysis (ABSA) involves three cascaded tasks: aspect (attribute) extraction, opinion extraction, and sentiment polarity classification. In the to‑restaurant scenario, four‑tuple extraction (dish, attribute, opinion, sentiment) is required for dish reviews, while three‑tuple extraction (attribute, opinion, sentiment) suffices for service and food‑safety reviews.

Technical Objectives

The goal is to build efficient, accurate models for four‑tuple and three‑tuple extraction, considering industrial constraints such as latency, resource consumption, and domain‑specific requirements.

Methodology

Pipeline Approach

The problem is decomposed into sequential modules: entity (dish) recognition, opinion extraction, aspect‑category classification, and sentiment classification. BERT+CRF is used for entity recognition; a QA‑style MRC model extracts opinions; multi‑task learning with shared BERT and attention layers handles aspect‑category and sentiment classification.

Joint Learning

To mitigate error propagation in the pipeline, a dual‑tower BERT model (Dual‑MRC) jointly extracts aspects, opinions, and sentiment in a single pass. Two queries are designed: one for dish entities, another for aspect‑opinion‑sentiment triples. The model outputs unified tags (e.g., B‑POS, I‑NEG) to capture both aspect and sentiment.

Few‑Shot and Prompt Techniques

For aspect‑category classification, prompt‑based P‑tuning and R‑Drop data augmentation are explored to improve performance under limited data.

Experiments and Results

Evaluation on Meituan’s annotated UGC data shows progressive improvements: BERT+CRF baseline F1 = 0.61, pipeline enhancements raise F1 to 0.65, and the Dual‑MRC model further boosts performance. Detailed ablation studies demonstrate the impact of learning‑rate scheduling, query design, and multi‑task sharing.

Business Applications

The extracted fine‑grained sentiments are applied to brand dashboards, dish information optimization, and review management tools, providing actionable insights for merchants and improving user experience.

Future Work

Planned directions include optimizing model efficiency, extending joint extraction to full four‑tuple modeling, and building a flexible, reusable framework for various fine‑grained sentiment scenarios.