Query Understanding Pipeline

The Query Understanding Pipeline (QU) is the front-end NLP intelligence hub of search engines and e-commerce portals that structurally decodes raw search intents within $<5\text{ms}$ before physical indexing: 1) Normalization (punctuation stripping, case folding, unicode normalization); 2) Spelling Correction (edit distance, phonetic pinyin mapping, and n-gram LM correction); 3) NER & Term Weighting (extracting brands, categories, attributes, and computing information entropy weights); 4) Synonym Expansion & Rewriting (e.g., expanding 'Apple phone' to 'iPhone'); 5) Intent & Category Classification (predicting transactional, navigational, or informational categories with confidence logits).

E-commerce search bars (Amazon/Taobao), general web search (Google), and vertical enterprise documentation search.

Raw search queries are ultra-short (2-3 words), noisy, typo-riddled, and ambiguous; QU translates unstructured queries into structured boolean query trees with category filters and weighted tokens, establishing the ceiling of retrieval accuracy.

Detail Term Weighting algorithms calculating importance weights between core entity nouns and modifier adjectives?

Design low-latency C++ inference acceleration for Transformer sequence-labeling NER models identifying e-commerce entities?

Explain the two-stage Query Spelling Correction architecture: Trie/phonetic candidate recall followed by n-gram/LM re-scoring?

How does synonym expansion prevent over-generalization via context-aware disambiguation (e.g. 'Apple' as fruit vs tech giant)?

How does Category Prediction enforce hard category filtering or soft boosting in e-commerce search to prevent irrelevant results?