Intelligent Customer Service at Meituan: NLP Techniques, System Architecture, and Real‑World Deployment

In this talk from the 2018 AI Pioneer Conference, Liu Xueliang introduces Meituan's intelligent customer service platform, outlining the historical evolution from voice call centers to modern AI‑driven multi‑channel systems.

The current architecture integrates several conversational agents: QABot for simple, single‑turn queries, TaskBot for complex, scenario‑based tasks, and ChatBot for casual chit‑chat, all supported by intent recognition, sentiment analysis, and knowledge‑base management.

Key NLP techniques include semantic recognition that maps diverse user expressions to standardized questions, a matching‑based approach using N‑gram features, and two high‑performing deep models—DSSM (a dual‑tower embedding model) and Seq2seq (used for similarity scoring). Knowledge‑graph embeddings and synonym extraction further enhance coverage.

TaskBot functions as an execution engine that triggers decision‑tree tasks, records node states, and invokes additional services (entity, sentiment, and semantic recognizers) to handle multi‑turn interactions such as payment issues.

ChatBot combines retrieval‑based and generative methods, balancing answer relevance with controllability for a professional support environment.

Operational metrics show that QABot resolves about 72,000 daily queries with a 92% offline accuracy and 83% online intelligent resolution rate, while TaskBot improves driver self‑service in the ride‑hailing domain.

The presentation concludes that intelligent, human‑machine collaborative customer service is essential for scaling Meituan's rapidly growing business, and future work will explore pre‑sale applications, intelligent marketing, and recommendation scenarios.