Semantic Equivalence of e-Commerce Queries
Source: https://queryunderstanding.com/semantic-equivalence-of-e-commerce-queries-8f3c2b1a9e4d Author: Daniel Tunkelang
Summary
How do you determine that two e-commerce queries are semantically equivalent? This is foundational for query clustering, synonym detection, and cache reuse. Presented work related to KDD 2023.
Problem Definition
Two queries Q1 and Q2 are semantically equivalent if they express the same shopping intent — i.e., a user would be equally satisfied by the result sets for either query.
This differs from lexical similarity (“laptop” vs “laptops”) and general semantic similarity (“big TV” vs “large television”).
Approach 1: Behavioral Aggregation
Queries that produce similar click distributions are semantically equivalent:
- Aggregate clicks by (query → item) pairs
- Represent each query as a weighted vector over clicked items
- Compute cosine similarity between query vectors
- Advantage: grounded in actual user behavior
- Challenge: data sparsity for rare queries
Approach 2: Sentence Transformers
Fine-tune a bi-encoder on pairs of (equivalent, non-equivalent) queries:
- Training signal: human-labeled pairs, or weak labels from behavioral similarity
- At inference: embed both queries, compare cosine similarity
- Advantage: generalizes to unseen queries
- Challenge: requires training data; may hallucinate equivalences
Hybrid Approach
Use behavioral similarity as weak supervision to train the sentence transformer:
- Compute behavioral similarity for high-traffic query pairs
- Label pairs above threshold as equivalent
- Fine-tune sentence transformer on these labels
- Apply transformer to low-traffic queries
Applications
- Query clustering: group equivalent queries for aggregate analytics
- Synonym expansion: rewrite rare queries to equivalent high-traffic forms
- Cache reuse: serve cached results for semantically equivalent queries
- A/B test bucketing: bucket equivalent queries together for cleaner experiments
Key Concepts
- Behavioral equivalence — click-distribution similarity as proxy for semantic equivalence
- Sentence transformers — bi-encoder fine-tuned on query pairs
- Weak supervision — using behavioral signal to generate training labels
- Query clustering — grouping equivalent queries