Intelligent QABot for 58.com: Classification and Retrieval Model Exploration

58.com, the largest domestic life‑service platform, created the "Bangbang" intelligent customer‑service (QABot) to automatically reply to user questions in micro‑chat, aiming to improve the connection between millions of B‑side merchants and C‑side users.

Because user messages are often short statements and many cannot be directly mapped to predefined categories, the answer rate of a pure classification system is low; therefore a two‑stage approach—classification plus retrieval—was adopted.

The classification pipeline collects micro‑chat logs, encodes user messages with BERT or word2vec into sentence vectors, clusters them using Bi‑KMeans, and defines a standard question for each cluster with expanded variations, enabling high‑frequency issue coverage.

Several industrial models were evaluated: FastText (fast baseline), LSTM+DSSM (adds sequential modeling), BERT (pre‑trained transformer), and the self‑developed SPTM framework (BERT‑style pre‑training with character‑level masking and Bi‑LSTM). Compared with FastText, LSTM+DSSM improved F1 by 6.07 %, BERT by 21.37 %, and SPTM by 20.79 %; SPTM also reduced inference latency to 11.74 ms versus 81 ms for BERT.

For retrieval, offline QA pairs ( ) are built after filtering with IDCNN entity recognizer and manual review. ALBert‑tiny encodes user messages into 312‑dimensional vectors, which are indexed with Faiss for nearest‑neighbor search; the top‑1 result is fetched from Redis during online inference.

Model fusion adopts a retrieval‑first strategy with the classification model as a fallback. This simple fusion raised the answer‑rate by 26.21 % without sacrificing accuracy, and future work will explore re‑ranking models and multi‑turn matching (e.g., DAM) to further improve semantic matching.

Overall, continuous algorithmic refinements of QABot have enhanced user experience and merchant conversion across multiple business lines, demonstrating the practical impact of AI‑driven customer service.