Improving Search Intent Recognition and Term Weighting with Deep Learning and Model Distillation at Ctrip

Background : In e‑commerce, search is the primary entry point for users, making query intent understanding and term weighting crucial for accurate results. Traditional pipelines rely on rule‑based matching, TF‑IDF, and statistical methods, which struggle with rare or ambiguous queries and cannot meet the millisecond‑level response times required in travel search.

Problem Analysis : To boost both accuracy and recall, the team introduced deep‑learning approaches, leveraging large‑scale pre‑trained language models (BERT) for richer semantic representations. However, the size of these models conflicted with strict latency constraints, prompting the use of model distillation and compression.

Intent Recognition : The core task is category identification for each token after segmentation. By feeding character‑level BERT embeddings into a pooling layer (max‑pooling performed best) and a parallel classifier, the system predicts the most probable category (e.g., city vs. hotel brand) with associated probabilities. Knowledge distillation transfers knowledge from a 12‑layer BERT teacher to a lightweight student model, preserving performance while reducing inference time to around 5 ms.

Term Weighting : User‑entered term importance is learned from click and suggestion data. A BERT‑based feature extractor converts each term’s characters into embeddings, which are then passed through a fully‑connected layer and softmax to predict term weights. The model is also distilled to meet a 2 ms latency target. Output weights (e.g., [0.34, 0.12, 0.53] for “Shanghai 的 Disney”) guide second‑stage recall and selective term dropping.

Future Outlook : The team plans to combine deep‑learning with traditional methods for head queries, explore larger domain‑specific pre‑trained models, and extend deep‑learning to other search components such as part‑of‑speech tagging and spelling correction, continually improving both performance and speed.