Phrase Matching

What is Phrase Matching?

Phrase matching requires adjacent, ordered tokens to appear together (or within a small window). It’s implemented with inverted indexes that store positions, phrase queries with slop, and/or bigram/phrase fields.

How It Works (quick)

• Index with positions: Store token positions per field for phrase/proximity queries.
• Query types: Exact phrase ("air max 270"), proximity with slop (allow 1–2 gaps), and bigram/trigram fields.
• Scoring: Combine with BM25/TF-IDF, field boosts (title > attributes > description).
• Normalization: Handle hyphens/diacritics; keep brand casing in exact fields.
• Blending: Try phrase matches first; fall back to unigrams/fuzzy when needed.

Why It Matters in E-commerce

• Model integrity: Lifts brand+model strings and prevents split-token noise.
• Natural queries: Captures common units (“trail running”, “phone case”).
• CTR gains: Better snippets/highlights increase click confidence.

Best Practices

• Add bigram fields for titles; use phrase queries with small slop.
• Protect SKU/MPN with exact keyword fields (no fuzz).
• Normalize variants (hyphen/space) into the same match family.
• Cap phrase boosts so poor items don’t outrank clearly better ones.
• Log missed phrases and expand analyzers where justified.

Challenges

• Language variation (“tee-shirt” vs “T-shirt”), tokenization across locales, and over-boosting exact but irrelevant phrases.

Examples

• "air max 270" in title outranks documents where tokens are scattered.
• "gore tex jacket" matches GORE-TEX jacket via normalized hyphenation.

Summary

Phrase matching encodes common word units and model names, improving precision and CTR. Use positions, small-slop phrase queries, and bigram fields—with careful normalization and sensible boosts.