Algorithm Evolution and Implementation of 58.com Intelligent QABot for Business Consultation

Background: 58.com, a leading Chinese lifestyle service platform, built an intelligent customer‑service system that combines automated Q&A with human agents to improve efficiency and user experience. The QABot provides three capabilities: business‑consultation Q&A, multi‑turn task dialogue, and casual chat.

QABot Question‑Answer Flow: When a user asks a business‑related question, the system matches it against a knowledge base. High‑confidence matches trigger a direct answer (unique response) or a TaskBot service; ambiguous matches return a list of candidate questions; unrelated queries are either treated as chit‑chat or rejected.

Knowledge‑Base Construction: The knowledge base is built in three steps—initially using manually edited FAQs, then expanding with labeled data from historical logs, and finally employing a semi‑automatic pipeline that uses a TextCNN classifier and Word2Vec embeddings followed by K‑means clustering to discover new standard questions.

Feature Engineering: Four iterations were performed. The first used character‑level features; the second added word segmentation with stop‑word and part‑of‑speech filtering; the third introduced keyword‑plus‑character features; the fourth combined word segmentation with selective POS replacement to balance performance and annotation effort.

Model Iterations: Three versions were deployed. Version 1 used a FastText multi‑class classifier with hierarchical softmax and threshold‑based decision rules. Version 2 introduced a two‑stage architecture combining a Bi‑LSTM binary reject model with FastText classification to improve recall and reduce over‑rejection. Version 3 replaced FastText with a Bi‑LSTM + DSSM deep semantic matching model and added a Bi‑LSTM intent recognizer to distinguish between answerable queries, service requests, list responses, and chit‑chat.

Evaluation: The system is measured by answer accuracy (percentage of correct answers) and machine‑intelligence resolution rate (proportion of interactions resolved without human hand‑off). Current figures are 92% accuracy and 86% resolution, saving the equivalent of 300–500 customer‑service staff.

Conclusion and Outlook: The QABot now achieves high performance, and future work includes incorporating contextual features, knowledge‑graph integration, and newer models such as BERT to further boost effectiveness.