Travel Search Technology and Innovations at Alibaba Feizhu

The talk introduces the evolution of travel‑scene search at Alibaba Feizhu, originally built to satisfy strong user demands for tickets, hotels, and attractions, and now transformed into a full‑text retrieval engine powered by AI.

Pig Search Background – Feizhu search consists of a global search entry and industry‑specific vertical searches. Users increasingly prefer the global entry because it reduces clicks, integrates traffic from Taobao, and offers the shortest interaction path.

Pig Search Framework – The pipeline starts with the Query Understanding (QP) service, which generates query intents and recall candidates. SP paging service calls the HA3 inverted index for recall, followed by coarse and weighted ranking, and finally LTP re‑ranking before results are displayed.

QP (Query Understanding & Recall Generation) – Challenges include strict latency (QP must occupy <10% of response time), traditional text relevance, travel‑specific needs such as LBS and POI understanding, and user‑feature personalization.

Query Tagging – Tags the query with destination and intent (e.g., "Beijing free‑travel" → destination: Beijing, intent: free‑travel). It involves three layers: offline tag‑library mining, online disambiguation using CRF/Tag algorithms, and applications like query correction and rewrite.

Product POI Mining – Extracts POI entities from product titles and unstructured HTML details using CRF++ and later a template‑based NER model, achieving >99% accuracy and >95% recall, greatly enriching POI features for retrieval.

Synonym Mining – Handles translation variants, bilingual terms, hierarchical relations, and typo corrections. A word2vec skip‑gram model trained on query‑title pairs generates candidate synonyms, combined with engineered features (edit distance, co‑occurrence, cosine similarity) and classified by LR/XGBoost, reaching 94% accuracy after manual review.

Correction – Uses a Hidden Markov Model to correct phonetic and shape‑similar character errors, supplemented by image‑based similarity (pixel‑wise comparison of font glyphs) and structural codes (Cangjie, Zhengma, Four‑Corner) to detect visually similar characters.

Recall Strategies

1. Classic Recall – Includes synonym mining, similar‑query rewrite, and POI mining. Query rewrite uses multi‑path candidates filtered by Learning‑to‑Rank (PS‑SMART), achieving 99% accuracy and reducing no‑result rate by 18%.

2. LBS Recall – Identifies POI in the query, obtains user coordinates, and restricts recall to nearby items, improving result relevance and achieving 95% accuracy.

3. Vector Recall – Embeds queries online and stores item embeddings in HA3 for ANN search. The model fuses query convolution features, item title convolution, and destination category via tensor‑fusion, uses large‑margin loss, and selects hard negatives randomly. This reduces no‑result rate by 32.7% and expands similar‑query coverage 1.7×.

4. Personalized Recall – Combines recommendation‑derived candidates (i2i, lbs2i, attribute2i) with text relevance filtering. User‑side features include profile attributes, recent interaction sequences, and attention over historical items; item‑side features include title, destination, and category. Tensor‑fusion merges these signals, improving personalization while keeping text relevance.

Model Optimizations – Text features are simplified to word‑vector concatenation to avoid over‑dominance in personalized scenarios. The final system integrates all recall lanes, with personalized recall feeding into ranking as a high‑rank feature.

Future Directions – Upgrade QP to "Query & User Planner" to incorporate richer user features, enhance explainability of search intents, and predict user destinations for ambiguous queries, aiming for more human‑readable intent understanding and proactive recommendation.

Overall, the presentation demonstrates how AI‑driven query understanding, sophisticated tagging, multi‑modal synonym mining, and diversified recall pipelines collectively improve travel search relevance, reduce no‑result cases, and lay groundwork for next‑generation personalized search at Alibaba Feizhu.