AI Techniques in Xiaomi Mobile Search: Text Relevance, Intent Recognition, and Click‑Model Ranking

Background: Recent advances in AI, especially deep learning, have been widely adopted in search and recommendation scenarios. Xiaomi's mobile search operates in two main contexts—MIUI's primary search entry and browser search suggestions—aiming to provide precise, convenient experiences by directly invoking relevant apps or services.

Typical Search Architecture: The system consists of a front‑end module, a query‑analysis module (handling intent, term weight, synonym expansion, and query rewriting), retrieval modules (keyword and vector‑based recall), followed by high‑level feature generation and re‑ranking (precision ranking) before returning results to the front end.

Text Relevance Calculation: The problem is defined as scoring the relevance between a query Q and a document title D, facing challenges like missing words, word order, and synonyms. Two core models are employed: a Gradient Boosted Decision Tree (GBDT) model using extensive engineered features, and a deep neural network (C‑DSSM) model that learns semantic representations from massive user‑behavior data. Model fusion leverages the strengths of both approaches, and the trade‑off between manually labeled high‑precision data and large‑scale behavior‑driven samples is discussed.

Intent Recognition and Term Weighting: Initial solutions relied on dictionary‑plus‑rule methods, which were later enhanced with Logistic Regression (LR) and Deep Neural Network (DNN) models. An FTRL‑optimized LR model enables online learning for dynamic intent changes. Features include N‑gram, length, position, and term‑weighting, with BiLSTM+CRF models providing state‑of‑the‑art performance for term‑weight estimation.

Click‑Through Ranking Models: To handle heterogeneous mobile search results (videos, apps, etc.), several click models are used: Position‑Based Model (PBM), Cascade Model, and Dynamic Bayesian Network (DBN) trained via EM algorithm. A multi‑model ensemble with additional bias features (e.g., presence of images, result size) further refines ranking for mobile scenarios.

Conclusion: The article summarizes the overall architecture of Xiaomi's mobile search, highlighting the integration of AI techniques across relevance computation, intent detection, and click‑model ranking, and provides insights into practical deployment and performance improvements.