Technical Exploration of Intelligent Dialogue Robots in Didi Ride-Hailing Scenarios

This presentation, delivered at the 2019 AI Science Frontier Conference, introduces the technical exploration of intelligent dialogue robots in Didi ride‑hailing scenarios, focusing on single‑turn QA, multi‑turn conversation, and the overall system architecture.

Single‑turn QA aims to accurately recognize user questions and provide appropriate answers, but faces challenges such as limited annotated data, numerous business lines (over ten), and diverse user expressions.

To address data scarcity, the team investigated data migration between similar business lines (e.g., fast‑car and premium‑car) and proposed a Multi‑Task Learning framework that shares a semantic model across lines while retaining line‑specific models for fine‑tuning. Various model backbones (CNN, LSTM, Transformer, BERT) were evaluated.

Experimental results show that adding Multi‑Task learning improves Top‑1 and Top‑3 accuracy, especially for BERT, while complex models may overfit when data is insufficient.

Beyond classification, a search + semantic matching + ranking pipeline was built for emotion‑support scenarios, leveraging the DUA model (Deep Utterance Aggregation) to incorporate dialogue context.

The multi‑turn dialogue framework consists of three layers:

Intent Recognition – implemented with BERT + Multi‑Task Learning.

Slot Extraction – a hybrid rule‑based and model approach (BiLSTM + CRF).

Dialogue Management – state tracking, policy selection (state‑machine and reinforcement‑learning based), and response generation.

When user intent is ambiguous, an Intelligent Questioning module uses knowledge‑graph queries and guided clarification (e.g., asking whether the order is real‑time or scheduled) to obtain precise information.

Chit‑chat handling (greetings, thanks, etc.) is currently based on classification and retrieval models, with ongoing exploration of generative approaches for more flexible responses.

The overall robot architecture routes user requests through a frontend that aggregates query and context, then a ranker selects among answer types (QA, task‑oriented, multi‑turn, chit‑chat, graph‑based) to deliver the final response.

The presented solutions support various Didi services (taxi, fast‑car, premium‑car), driver assistants, and internationalized customer service, illustrating a comprehensive AI‑driven intelligent客服 system.