58.com Dialogue Robot Application Practice in Local Life Service: Architecture, Core Capabilities, and Future Outlook

58.com is a large local‑service platform connecting millions of B‑side merchants with C‑side users; the company leverages a dialogue robot to improve the efficiency and revenue of merchants in the yellow‑pages (local life) scenario.

In the AI‑driven mode, all user inquiries are first handled by an automatic Q&A robot; when the robot cannot answer, the conversation is transferred to a human agent based on configurable time windows.

Core Capabilities

QABot addresses user question‑answering by building a business‑line knowledge base through clustering (e.g., KMeans) and manual annotation, training a matching model (FastText, DSSM, LSTM, BERT, SPTM) to map queries to standard question IDs, and generating answers from a structured answer pool. The model achieves F1 > 0.8 after multiple iterations.

TaskBot guides users to provide specific "opportunity" slots (e.g., phone number) using an intent‑slot framework. It combines QABot for standard queries and TaskBot for multi‑turn slot‑filling, achieving higher conversion rates in real‑world deployments.

Intelligent Lead (Smart Lead) Generation uses the dialogue manager’s state and timed prompts to proactively ask users for phone numbers, mimicking human agents’ follow‑up behavior, which improves lead retention by 14%.

Opportunity Mining extracts phone, location, time, and other business attributes from both robot‑handled and human‑handled conversations using IDCNN+CRF models, automatically feeding them into the opportunity management system for timely merchant follow‑up.

The overall dialogue management service is built on a state‑machine framework where each node represents a combination of intents and slots, stored in WRedis, and actions are selected to generate responses via templates.

Future work will focus on algorithmic innovation such as reinforcement learning and further pre‑training techniques to continuously improve the dialogue robot’s effectiveness.