Real-World End-to-End Product Entity Resolution

Dimitrios Karapiperis, George Papadakis, Manolis KoubarakisarXiv 2026 · arXiv

TL;DR

This paper provides a comprehensive end-to-end evaluation of zero-shot Product Entity Resolution (PER) pipelines, moving beyond simple matching to look at the entire lifecycle of candidate generation and verification. By introducing the new RETAILERS benchmark and testing everything from BM25 to LLMs, the authors reveal that while semantic models are powerful, they often stumble on fine-grained numeric details where traditional lexical methods excel—making hybrid architectures the current "gold standard" for real-world e-commerce.

What Problem Does This Paper Address?

Product Entity Resolution (PER) is the art of determining if two different online listings—say, a "Sony PS5 Digital Ed." on one site and a "PlayStation Five w/ Controller" on another—refer to the exact same physical product.

While standard Entity Resolution (like matching names in a census) is well-studied, e-commerce is a "hostile environment." Listings are messy, filled with promotional jargon ("Ships free!"), and highly sensitive to tiny alphanumeric changes. A single character difference (e.g., "v1" vs "v2" or "128GB" vs "256GB") defines an entirely different product.

Most current state-of-the-art systems rely on heavy fine-tuning on labeled data. However, in the real world, products change daily, and labeling thousands of pairs for every new category is impossible. This paper tackles the zero-shot challenge: how do we build robust, scalable pipelines that work out-of-the-box without task-specific training?

Methods

The authors evaluate a two-stage architecture: Blocking (quickly pruning millions of items to a few likely candidates) and Matching (rigorously verifying those candidates).

1. Blocking Strategies

  • Lexical Cascade: Uses BM25 (keyword matching). Fast and precise with numbers but fails on synonyms (e.g., "TV" vs "Television").
  • Semantic Cascade: Uses Bi-encoders to map products into a vector space. Great for synonyms but often "blurs" distinct products that look similar (the "low-pass filter" effect).
  • Hybrid Fusion: Combines both using Reciprocal Rank Fusion (RRF) to get the best of both worlds.
  • Generative-Lexical: Uses an LLM offline to "clean" and normalize titles into a canonical format before using standard lexical matching.

2. Matching Engines

  • Cross-Encoders: Deep Transformer models that look at both strings simultaneously.
  • Tensor Search (ColBERT): A "late-interaction" model that is faster than cross-encoders but less precise.
  • LLMs (Gemma & GPT): Using Large Language Models as zero-shot pairwise verifiers.

Datasets

The study uses four established benchmarks (ABT-BUY, AMAZON-GOOGLE, AMAZON-WALMART, WDC-Products) and introduces a new one:

  • RETAILERS (C and D variants): A novel dataset of fast-moving consumer goods. It features cross-lingual noise (mixing local languages with English brand names), inconsistent units (1L vs 1lt), and "brand collisions" where different manufacturers have very similar product names.

Key Insights and Ideas

  • Complementary Failure Modes: Lexical and semantic retrieval are "mirror images" of each other. Lexical methods fail when words change but meaning stays the same; semantic methods fail when words stay similar but a tiny numeric change alters the meaning.
  • The Memory Ceiling: Surprisingly, the primary bottleneck for scaling dense vector search isn't GPU memory—it's host CPU RAM. Storing the spatial adjacency lists for HNSW (a common vector index) requires massive amounts of uncompressed RAM.
  • The "Low-Pass Filter" of Embeddings: Dense embeddings are great at capturing "vibe" but terrible at "specs." They often cluster "iPhone 13 128GB" and "iPhone 13 256GB" so closely that the blocking phase can't tell them apart.
  • LLM Divergence: There is a massive performance gap between local and remote LLMs. Local models (like Gemma) struggle with long, noisy descriptions, while remote models (like GPT) leverage vast pre-trained knowledge to resolve complex promotional noise.

Key Results

  • Hybrid is King for Recall: The Hybrid Fusion approach consistently achieved the highest blocking recall (e.g., 0.958 on RETAILERS-C), proving that you shouldn't ditch BM25 just because you have AI.
  • Cross-Encoders are the "Practical Default": For the matching stage, cross-encoders provided the best balance of accuracy and speed. Tensor search (ColBERT) was faster but suffered significant precision loss.
  • The LLM Speed Tax: While LLMs can achieve peak F1 scores (up to 0.872 on ABT-BUY), they are incredibly slow. A local Gemma-based pipeline took 1,300 seconds to process a task that a cross-encoder finished in 15 seconds.
  • Generative Normalization Works: Using an LLM to pre-process and "canonicalize" titles (Generative-Lexical Cascade) offered a 700x speedup over using the LLM for pairwise matching, with only a modest hit to accuracy.

Limitations and Future Work

  • Hallucinations: The Generative-Lexical approach risks the LLM "hallucinating" specifications during the normalization phase, which then become permanently encoded in the search index.
  • Scalability: Pairwise LLM matching remains computationally and financially prohibitive for massive datasets like WDC-Products (millions of listings).
  • Future Directions: The authors suggest moving toward disk-aware ANN indexes (like DiskANN) to solve the CPU RAM bottleneck and developing adaptive blocking that can decide on-the-fly whether to use lexical or semantic search for a specific query.

Summary updated

Abstract

Product Entity Resolution (PER) in e-commerce is challenging due to unstructured text, formatting inconsistencies, promotional noise, and fine-grained product differences. We present an end-to-end evaluation of four zero-shot PER pipelines---Lexical Cascade, Semantic Cascade, Hybrid Fusion Cascade, and Generative-Lexical Cascade---across established benchmarks and introduce two real-world datasets characterized by cross-lingual noise and dense brand collisions. We evaluate lexical and dense blocking with cross-encoder, tensor-search, and Large Language Model verification, considering accuracy, runtime, and memory consumption.
BibTeX
@misc{karapiperis2026real,
  author = {Dimitrios Karapiperis and George Papadakis and Manolis Koubarakis},
  title = {{Real-World End-to-End Product Entity Resolution}},
  year = {2026},
  eprint = {2611.19418},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19418},
}
Figure