Model Iteration and Architecture of the BangBang Intelligent Customer Service QABot

The BangBang intelligent customer service, developed by 58.com for real‑estate agents, aims to deliver accurate, fast, and efficient Q&A services, reducing manual客服 costs and improving operational efficiency.

The system’s overall technical architecture comprises a main service, ABTest service, annotation system, knowledge discovery module, and intent‑recognition component, all configurable for rapid scenario integration.

Core capabilities include intent recognition that determines whether to provide a unique answer, a list of possible answers, self‑service actions (e.g., displaying violating posts or creating tickets), or reject the query as unrelated chatter.

The QABot relies on a knowledge base built from standard and expanded question pairs; matching is performed via keyword matching and a multi‑class classification model where each knowledge‑base entry represents a class.

Model iteration began with FastText (dim=256, wordNgrams=3, bucket=300000, lr=1e‑3, epoch=200), achieving 81.45% accuracy, 95.34% recall, and an F1 score of 0.8785.

Switching to TextCNN (embedding=256, filters=[2,3,3,4,3]×64, batch=128, epoch=80, lr=1e‑4) raised accuracy to 90.13%, recall to 92.89%, and F1 to 0.9149.

A two‑layer bidirectional LSTM model (hidden_layers=64, rnn_dim=128, max length=80, lr=1e‑4, batch=128, 60 epochs) further improved performance to 93.94% accuracy, 92.51% recall, and an F1 of 0.9322.

Finally, model fusion combined a Bi‑LSTM for answer‑type prediction (Model A) and another Bi‑LSTM for intent prediction (Model B), using threshold‑based strength levels; this approach boosted metrics to 94.89% accuracy, 93.33% recall, and an F1 of 0.9410.

In summary, the QABot’s metrics have stabilized around 95%; future work will explore stronger pre‑trained models such as BERT and further knowledge‑base enhancements to continue improving answer quality and self‑service coverage.