Meituan's End-to-End Sentiment Analysis Technology and the ASAP Dataset

Background

In May 2021, Meituan's NLP Center released ASAP, the largest real‑world Chinese attribute‑level sentiment analysis dataset to date. The dataset and its accompanying paper were accepted at NAACL 2021 and are part of the open‑source data initiative “Qianyan”. The article reviews the evolution of Meituan’s sentiment analysis techniques and their applications in various business scenarios.

Sentiment Analysis Overview

Sentiment analysis aims to identify positive (praise), negative (criticism), or neutral sentiment in text. Depending on granularity, it can be performed at the document/sentence level or at the attribute (aspect) level, the latter often referred to as Aspect‑Based Sentiment Analysis (ABSA).

Document/Sentence‑Level Sentiment Analysis

The task predicts an overall sentiment intensity or class for a whole document or sentence. Meituan adopted a regression formulation (continuous score in [0,1]) with seven discrete sentiment grades. Early models such as TextCNN, Att‑BLSTM, and BERT were evaluated; the Att‑BLSTM architecture was later upgraded with a self‑developed MT‑BERT encoder, yielding significant performance gains (see Figure 6).

These models are deployed in real‑time content safety, recommendation, and business intelligence pipelines, filtering negative content and monitoring sentiment trends for merchants.

Attribute‑Level Sentiment Analysis (ABSA)

To capture fine‑grained opinions, Meituan defined a two‑level attribute taxonomy for the restaurant domain (e.g., taste, environment, price, service, location) and expanded it to 18 sub‑attributes. Each comment is annotated with one of four labels per attribute: [not mentioned, negative, neutral, positive]. The ASAP dataset contains far more comments and attributes than the SemEval RESTAURANT dataset, presenting a challenging multi‑task learning problem.

The final architecture follows a shared encoder (Bi‑GRU → BERT → MT‑BERT) and task‑specific heads for each attribute (Figure 10). Multi‑task learning mitigates data imbalance and enables knowledge transfer across attributes.

Opinion Triplet Extraction

Triplet extraction jointly predicts aspect, opinion, and sentiment (Aspect‑Opinion‑Sentiment). Meituan uses a multi‑task model comprising a pretrained language encoder, aspect tagger, opinion tagger, and a biaffine sentiment parser (Figure 14). The system achieves strong EM‑F1 and Fuzzy‑F1 scores, and the extracted triplets are used to build an aspect‑based sentiment knowledge base for recommendation copy and merchant operation insights.

Serviceization

Two services were built: an online real‑time prediction API for latency‑sensitive scenarios (e.g., dialogue systems) and an offline batch prediction pipeline for large‑scale content processing. Both services are powered by Meituan’s ML platform and now serve more than ten internal business lines.

Summary and Outlook

The paper summarizes Meituan’s sentiment analysis pipeline—from document‑level to attribute‑level and triplet extraction—highlighting the release of ASAP and continuous model upgrades with MT‑BERT. Future work includes transfer learning, few‑shot learning, and automatic attribute discovery to accelerate deployment in new domains.

References

Key references include the ASAP dataset paper (NAACL 2021), foundational works on ABSA, BERT, Att‑BLSTM, and recent multi‑task triplet extraction studies.