Ecommerce ops · Production

Delivery Hero builds semantic product matching using retrieval-rerank and hard negative sampling

The problem

Delivery Hero needed an algorithm to match product titles across its own catalog and competitors' catalogs for pricing strategy and assortment gap analysis, and to detect internal duplicate items.

First attempt

Lexical matching could not recognize the same product described with different words — unit abbreviations, misspellings, and missing words all produced false non-matches.

Workflow diagram · grounded in source
1
Product title query submitted
trigger
“our goal is to take a given product title, referred to as p, and find a corresponding product title within an unordered set, labelled as S={t_1, t_2, …, t_n}”
2
Lexical candidate retrieval
ai_action
“In our specific case, we opted for the first approach, leveraging Lexical Matching. This choice was driven by its cost-effectiveness and efficiency, making it well-suited for rapidly processing a large number of documents to extract a ma…”
3
Cross-encoder reranking
ai_action
“we employed a transformer-based cross-encoder model. Unlike an encoder-only model, which processes inputs independently and generates embeddings for each, the cross-encoder examines pairs of inputs together. This joint processing allows …”
4
Hard negative fine-tuning
feedback_loop
“Once these hard negative samples were identified, we proceeded to fine-tune our cross-encoder models using the hard negative samples”
Reported outcome

The Retrieval-Rerank approach augmented by hard negative sampling enables effective identification of similar products and assortment gap management while balancing computational efficiency and accuracy.

Reported metrics
Matching accuracysignificantly higher accuracy
Product matching and assortment gap capabilityeffectively identify similar products, manage assortment gaps
Reported stack
SBERTLucene
Source
https://tech.deliveryhero.com/semantic-product-matching/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The Retrieval-Rerank approach augmented by hard negative sampling enables effective identification of similar products and assortment gap management while balancing computational efficiency and accuracy.

What tools did this team use?

SBERT, Lucene.

What results were reported?

Matching accuracy: significantly higher accuracy; Product matching and assortment gap capability: effectively identify similar products, manage assortment gaps (source-reported, not independently verified).

What failed first in this deployment?

Lexical matching could not recognize the same product described with different words — unit abbreviations, misspellings, and missing words all produced false non-matches.

How is this ecommerce ops AI workflow structured?

Product title query submitted → Lexical candidate retrieval → Cross-encoder reranking → Hard negative fine-tuning.

Related ecommerce ops cases