Multi-turn Response Triggering Model (MRTM) for Intelligent Customer Service Chatbots

This article offers a comprehensive review of Didi's research paper titled "Towards Building an Intelligent Chatbot for Customer Service: Learning to Respond at the Appropriate Time". The paper introduces a Multi‑turn Response Triggering Model (MRTM) that enables a chatbot to reply only at appropriate moments while listening at other times, thereby improving dialogue fluency and response quality.

Research Background In recent years, intelligent dialogue robots have become increasingly prevalent in customer service. Traditional turn‑by‑turn interaction often leads to inappropriate or repetitive replies, especially when users send short, fragmented, or repeated queries on mobile devices. Existing studies on optimal reply timing are scarce; for example, Google’s Smart Reply uses a binary triggering model trained on annotated email data, which does not translate well to the semi‑open, multi‑turn nature of customer‑service conversations.

Problem Challenges The paper identifies three main challenges: (1) the need for multi‑turn modeling so that identical user queries may receive different responses depending on context; (2) high precision and recall requirements, as both over‑replying and under‑replying degrade user experience; (3) practical constraints such as agents handling multiple customers simultaneously, leading to delayed human replies and making supervised learning on reply‑timing signals infeasible.

Proposed Solution MRTM is built on a self‑supervised learning framework. It introduces a Multi‑turn Response Selection auxiliary task: given a dialogue context and candidate replies, the model selects the correct reply, thereby learning semantic matching between context and response. An asymmetric self‑attention mechanism is applied within sliding windows of local context, using the last sentence as the key to compute attention scores. This design emphasizes sentences that are closer to the agent’s reply, assuming they are more likely to require a response, while suppressing less relevant utterances. The model aggregates representations of the final local‑context windows and combines them with all agent replies to capture temporal dependencies. Both BiLSTM and BERT encoders are explored, with BiLSTM chosen for online deployment due to efficiency.

Experiments and Results Four baselines are compared: two rule‑based methods (Active Trigger based on Longest Utterance – ATLU, Passive Trigger based on Shortest Utterance – PTSU) and two supervised learning methods (Supervised Single‑turn Triggering – SST, Supervised Multi‑turn Triggering – SMT). Experiments on Didi and JD customer‑service datasets show that MRTM significantly outperforms all baselines in both accuracy and recall. Rule‑based ATLU achieves high accuracy but low recall, while PTSU performs poorly on both metrics. Among supervised methods, SMT benefits from stronger multi‑turn modeling than SST. The BERT‑based MRTM yields modest gains over the BiLSTM version, but the latter is retained for production because of lower latency.

Conclusion The self‑supervised MRTM effectively learns when a chatbot should respond, addressing the challenges of multi‑turn, semi‑open dialogues in customer service. Its integration into Didi’s intelligent客服 system demonstrates practical value for drivers, passengers, and support agents.