How JD’s Heuristic QA Boosts Smart Customer Service with AI

1. Background

In recent years, the increasing use of intelligent customer‑service robots on e‑commerce platforms has made user satisfaction and problem‑resolution rate the key metrics for service quality. User experience depends on communication convenience and answer accuracy, prompting the proposal of a heuristic question‑answering framework.

2. Deployment

The JD intelligent‑service heuristic QA consists of three modules: pre‑consultation question prediction, in‑consultation associative input, and post‑consultation related recommendation. About one‑third of messages can now be completed by clicking candidate queries, greatly improving communication convenience and boosting overall answer accuracy, problem‑resolution rate and user satisfaction.

3. Algorithm Technology

3.1 Pre‑consultation Question Prediction

Before a user initiates a chat, the robot predicts the user’s intent and pushes several candidate queries to the front‑end, reducing input effort and normalising expressions.

3.1.1 Prediction Engine

The engine is built on massive business data and offers fast integration of various prediction needs, data accumulation, a unified prediction model, and self‑learning capabilities.

Data layer stores massive business and anonymised user data; storage layer uses Flink and HBase; algorithm layer employs DeepCTR‑based models such as DeepFM and DCN; service layer provides real‑time prediction; business layer includes online and voice bots and human agents.

3.1.2 Cold and Hot Start

Cold start collects external system data, user intent and service‑order information to train a prediction model before it can serve requests. Hot start directly uses cached data and the trained model to answer queries in real time.

3.1.3 Prediction Model

Using 532 k samples across 120 classes, DCN achieved 40.73 % accuracy, outperforming DeepFM, Wide&Deep and LSTM‑augmented models.

3.1.4 Self‑Iterating Model

The model monitors intent distribution shifts (e.g., during large promotions) and automatically triggers retraining when deviation exceeds a threshold, ensuring up‑to‑date predictions.

3.2 In‑consultation Associative Input

Also known as query auto‑completion, the system predicts full queries from user prefixes, applies sensitive‑word filtering, traditional‑simplified conversion and text correction, and ranks candidates with a Learning‑to‑Rank model using intent, text and statistical features.

3.3 Post‑consultation Related Recommendation

After answering a question, the system recommends the next likely user queries, combining AI‑driven mining with manual configuration to achieve a 1+1>2 effect.

Offline, mining models extract candidate related questions; online, a ranking model (using text, intent and business features) re‑orders them before display.

Experiments show the XGBoost‑fused model reaches a top‑3 accuracy of 97.9 %.

4. Future Improvements

Future work includes unified heuristic QA prediction, multi‑turn dialogue enhancement, and fully automatic deep‑semantic multi‑turn QA without manual configuration.