Architecture and Evolution of the 58 Intelligent Q&A Chatbot System

Background

58.com, a leading Chinese lifestyle service platform, generates massive Q&A data from interactions between users, merchants, and customer service agents. To improve communication efficiency, the 58 AI Lab built an intelligent Q&A robot that combines natural language processing, deep learning, semantic understanding, and intelligent recommendation to accurately recognize user intent and provide automated answers.

Overall Architecture

The system adopts a three‑layer design: an access layer, a logic layer, and a data layer. The access layer handles template, gateway, Q&A management, and business‑opportunity services. The logic layer, built on micro‑services, includes NLU, rule‑matching, retrieval, prediction, ranking, and dialogue‑management services. The data layer stores Q&A, operational, and knowledge‑base data, supporting model training and real‑time updates via annotation and integration platforms.

QABot

QABot provides business‑specific question answering by leveraging deep‑learning models, rule‑based matching, and retrieval. User queries are processed through NLU, rule matching, retrieval (BM25), and prediction (BI‑LSTM+DSSM) services, then merged and ranked to produce a single answer, a list, a chit‑chat response, or a fallback.

To meet growing traffic, the original serial processing pipeline was parallelized: rule matching, retrieval, and prediction run concurrently, and their results are fused according to priority, reducing overall latency to the longest‑running sub‑service.

Performance optimizations also include multi‑level caching (in‑memory loading, WMB message‑bus synchronization, and LRU eviction) to ensure millisecond‑level data freshness for knowledge‑base updates.

TaskBot

TaskBot enables multi‑turn dialogues for complex business scenarios and proactive business‑opportunity guidance. Its architecture consists of a data layer (NLU metadata, multi‑turn Q&A store, dialogue scripts) and a logic layer (dialogue state tracking, policy selection, and natural language generation). The system integrates NLU, knowledge‑graph disambiguation, and configurable strategy boxes to allow plug‑and‑play workflow adjustments.

Answer Recommendation

The recommendation module offers input suggestions, "you may also ask", and scene‑based ranking to guide users and reduce input errors. Initially it used BM25 for recall; later it incorporated a full recommendation pipeline (recall → ranking) with algorithms such as TF‑IDF, DSSM, XGBoost, and DeepFM, all experimentable via the ABTest platform.

Dynamic Strategy Adjustment

All services are wrapped in a plug‑in strategy box architecture, allowing AB testing of multiple models and policies simultaneously. This enables rapid iteration, continuous performance improvement, and easy customization for different business lines and scenarios.

Conclusion

The iterative evolution of the intelligent Q&A robot exemplifies a "small steps, fast iteration" development philosophy. By progressively launching QABot, TaskBot, and recommendation services and then optimizing performance, caching, and strategy configurability, the system now achieves near‑human accuracy and conversion rates in customer‑service and merchant‑assistant scenarios.