Future Intent Prediction for Chatbots: Architecture, Techniques, and Evaluation

JD.com’s JIMI is an in‑house chatbot built on natural language processing, deep neural networks, and machine‑learning techniques, serving billions of user queries with over 90% answer accuracy and high satisfaction.

The existing architecture consists of three main modules: an algorithm layer handling error correction, tokenization, entity recognition, knowledge graph and lexical analysis; an engineering layer that routes questions based on business logic; and a data layer that aggregates, cleans, and visualizes customer‑service knowledge.

Traditional chatbots only answer the current user query, lacking any ability to anticipate the next user intent, which leads to repetitive input and poor experience.

The proposed solution adds future‑intent prediction in two scenarios: (1) before the user starts a conversation, by clustering historical questions, filtering them with a logistic‑regression model, generating word embeddings via word2vec, and selecting the top‑20 clusters as standard prompts; (2) in real‑time during a conversation, by continuously predicting the next intent and displaying the top‑5 likely questions for one‑click answers.

The prediction pipeline includes preprocessing (validating input length), model inference (computing probability of next question and applying a threshold), and data logging for offline model tuning.

Training data are constructed by concatenating the first N‑1 user utterances as a sample and labeling it with the N‑th utterance’s category, using manually curated taxonomy; only the most frequent ten categories are kept for prediction.

A single‑layer CNN model with 100‑dimensional word vectors, 30‑character truncation/padding, and a 3×50 convolution kernel is employed, achieving click‑through rates of 71.2% and accuracy of 78.2%, outperforming baseline and LSTM‑based approaches.

Online experiments confirm that future‑intent prediction reduces consultation time, improves user satisfaction, and makes the chatbot not only understand the current question but also anticipate the user’s next need.