Meituan Intelligent Customer Service: Technical Practices and Core Technologies

Meituan, a large lifestyle e‑commerce platform, faces massive user‑service interactions across pre‑sale, in‑sale, post‑sale, and internal office scenarios, prompting the development of an intelligent customer service system to handle simple queries automatically and route complex issues to human agents.

The system’s six core capabilities include problem recommendation, problem understanding, dialogue management, answer supply, script recommendation, and conversation summarization, each designed to improve efficiency, reduce cost, and enhance user experience.

Dialogue interaction technology is central, with tasks classified as chit‑chat, task‑oriented, and QA. Implementation methods span retrieval‑based, generation‑based, and task‑oriented approaches, leveraging models such as BERT, RoBERTa, and multi‑task multi‑field architectures to achieve near‑90% intent recognition accuracy.

Multi‑turn dialogue combines intent matching, task flow execution, and answer generation, supported by a Bayes‑Network‑based TaskFlow engine for fuzzy matching and a visual TaskFlow editor for business teams to design processes without coding.

Human‑assisted features include script recommendation—using historical "N+1" QA pairs and BERT‑based ranking to suggest replies—and conversation summarization—employing classification, BERT‑Sum extraction, and PEGASUS generation models to produce concise summaries for ticket creation and analytics.

All capabilities are unified in the Moses dialogue platform, which offers layered solutions (application scenarios, solution packs, dialogue abilities, and platform functions) and supports varying business maturity levels, from simple intent‑based bots to advanced API‑driven task flows.

Future work focuses on end‑to‑end dialogue modeling, data‑driven robustness, and empathetic response generation to better handle user emotions in customer service contexts.