🤖 machine learning

Domain-Specific Text Embedding Models for Entity Resolution

This paper demonstrates that fine-tuning general-purpose text embedding models with domain-specific triplet training significantly improves their ability to distinguish between identical and highly similar non-matching entity records, thereby enhancing performance in entity resolution and duplicate record retrieval tasks.

Original authors: Khajesh Sapram, Srivardhani Raju, Kishore Konda

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Khajesh Sapram, Srivardhani Raju, Kishore Konda

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, humming archives of the modern world, information is rarely neat. A single business or person might be recorded in a dozen different ways: "Inc." instead of "Incorporated," a street address with a missing zip code, or a name spelled slightly differently in a new system. For computers to make sense of this chaos, they rely on a technology called text embeddings. Think of these as a way to translate words into a map of numbers, where similar meanings sit close together. If you ask a computer to find documents about "cats," it looks for words that live near the concept of a cat in this mathematical space. This works beautifully for general questions, but it stumbles when the goal is to find the exact same entity, even if the text has changed. A computer might think two different coffee shops on the same street are the same because they share a name, or it might fail to realize that "Ltd." and "Limited" refer to the very same company. This is the problem of entity resolution: distinguishing between two records that look alike but are different, and two records that look different but are the same.

A team of researchers set out to solve this specific puzzle by teaching computers to care less about general meaning and more about identity. They started with two powerful, pre-existing computer models that were already very good at understanding general language. These models had been trained on massive amounts of text to understand that "happy" and "joyful" are close neighbors. However, the researchers knew these models were not designed to tell the difference between a company in London and a different company with the same name in Manchester. To fix this, they created a synthetic dataset—a carefully constructed collection of fake business and person records. They built these records to mimic the messy reality of real-world data, introducing realistic variations like abbreviations, punctuation changes, and reordered address details. Crucially, they also created "hard" negative examples: records that were textually very similar but belonged to different entities, such as two people with the same name living in different cities.

The researchers then used a teaching method called triplet learning to reshape how these models saw the world. In this process, the computer was shown three items at once: a reference record, a variation of that same record, and a different record that looked very similar. The goal was simple but demanding: the computer had to learn to pull the reference and its variation closer together, while pushing the different, look-alike record further away. It was a lesson in nuance, forcing the model to ignore superficial changes like a missing comma or a shortened street name, while becoming hyper-sensitive to the one detail that actually changed the identity, such as a different branch location. They tested this approach on two widely used models, measuring how well they could separate the true matches from the tricky look-alikes.

The results were striking. Before this targeted training, the models struggled to make a clear distinction. When faced with the strictest test, where the difference between a match and a non-match had to be very obvious, the best-performing pre-trained model only succeeded about 15 percent of the time. It often couldn't tell the difference between a company in London and a different company in Manchester if they shared a name. After the domain-specific training, the picture changed completely. The same model, now fine-tuned on the specific task of identity resolution, jumped to a success rate of over 92 percent. The other model tested showed similar dramatic improvement, rising from roughly 38 percent to over 83 percent. The training didn't just make the models slightly better; it fundamentally reshaped their internal map, creating a space where records of the same entity huddled together, while records of different entities, no matter how similar they looked on the surface, were pushed apart.

The study suggests that the key to solving these identity puzzles is not to build a new computer from scratch, but to teach existing ones to pay attention to the right details. By focusing on the specific differences that matter in business and personal records, the researchers demonstrated that general-purpose tools can be adapted to handle the high-stakes precision required for data quality. The findings indicate that with the right kind of training, computers can learn to ignore the noise of formatting and spelling while locking onto the signal of true identity. This approach offers a practical path forward for organizations drowning in duplicate records, allowing them to clean their data and find the right connections without relying on rigid, manual rules. The work points toward a future where information retrieval systems are not just smart about meaning, but are also sharp enough to know exactly who and what they are talking about.

Technical Summary: Domain-Specific Text Embedding Models for Entity Resolution

Problem Statement
General-purpose text embedding models are optimized to capture broad semantic similarity but often fail in Entity Resolution (ER) and duplicate record retrieval. In ER, the objective is not merely to find semantically related text, but to distinguish whether two records refer to the same real-world entity (e.g., a specific business or person). General models struggle because they may assign high similarity scores to different entities that share names or locations, while failing to sufficiently separate genuine duplicates that contain minor textual variations (e.g., abbreviations, formatting differences, or transliterations). Traditional rule-based or edit-distance methods are interpretable but require extensive domain expertise and struggle with complex, simultaneous variations. Conversely, standard dense retrieval models lack the sensitivity to fine-grained identity distinctions required for reliable ER.

Methodology
The paper proposes adapting general-purpose sentence embedding models for identity-sensitive retrieval through domain-specific triplet fine-tuning.

Dataset Construction: A synthetic dataset of business and person records was created to provide controlled, realistic variations without exposing confidential information. The dataset includes:
- Anchors: Reference records.
- Positives: Variations of the anchor representing the same entity (introducing abbreviations, punctuation changes, reordering, or omissions).
- Negatives: Different entities that are textually similar to the anchor but contain meaningful identity or location conflicts (e.g., same company name, different address).
- The dataset emphasizes "hard" and "moderately difficult" negatives to prevent the model from relying on superficial cues.
Models Evaluated: Two widely used pretrained models were selected as baselines: sentence-transformers/all-MiniLM-L6-v2 and BAAI/bge-base-en-v1.5.
Training Strategy: Both models were fine-tuned using triplet loss with cosine distance. The objective function minimizes the distance between the anchor and positive while maximizing the distance between the anchor and negative, specifically targeting a margin $m$ .
Evaluation Protocol: Performance was measured on a held-out validation set using a margin-based similarity evaluation. Instead of simple ranking, the study measured the "Score Gap" ( $S_{positive} - S_{negative}$ ) and calculated pass rates at specific margins (0.10, 0.20, and 0.30). A higher margin indicates a stronger, more reliable separation between matching and non-matching records.

Key Contributions

Triplet Construction Methodology: A framework for generating synthetic business and person records that capture realistic identity-preserving variations and challenging identity-conflicting negatives.
Evaluation Protocol: A margin-based metric designed to quantify the quality of separation in dense embedding spaces for identity-sensitive tasks, moving beyond simple ranking accuracy.
Empirical Adaptation Study: An analysis of how domain-specific supervision reshapes the embedding spaces of two distinct, widely adopted sentence embedding models.
Evidence of Efficacy: Demonstration that targeted fine-tuning significantly improves the retrieval of duplicate entity records while increasing separation from highly similar non-matching candidates.

Results
The study found that general-purpose models, while capable of ranking matches above non-matches, often produce small score gaps, making threshold selection difficult.

Pretrained Baselines: At the strictest margin (0.30), the pretrained all-MiniLM-L6-v2 achieved a 37.85% pass rate, and BGE Base EN v1.5 achieved 15.25%.
Fine-Tuned Performance: After domain-specific triplet fine-tuning, both models showed substantial improvements.
- BGE Base EN v1.5 improved from 15.25% to 92.70% at margin 0.30.
- all-MiniLM-L6-v2 improved from 37.85% to 83.10% at margin 0.30.
Comparative Analysis: While all-MiniLM-L6-v2 had a stronger initial baseline, BGE Base EN v1.5 demonstrated a greater capacity for adaptation, outperforming the other model by approximately 9.6 percentage points at the 0.30 margin after training.

Significance and Claims
The paper claims that domain-specific triplet training effectively reshapes general-purpose embedding spaces to be better suited for entity resolution. The findings suggest that targeted fine-tuning is a practical approach for improving data quality management and information retrieval applications where identity preservation is critical.

The authors emphasize that these results bridge the gap between broad semantic similarity and identity-sensitive retrieval. They note that while the synthetic dataset allows for precise control, real-world applications may face additional noise and distributional shifts. The study concludes that future work should explore context-conditioned embedding models, where retrieval context (e.g., "ignore legal entity designations but prioritize address changes") is part of the model input, potentially moving beyond static, task-specific embeddings toward more adaptive retrieval systems.

Technical Summary: Domain-Specific Text Embedding Models for Entity Resolution

More like this