PaperlayerView exact revision on arXiv

Source-linked AI summary

Domain-Specific Text Embedding Models for Entity Resolution

Khajesh Sapram, Srivardhani Raju, Kishore Konda

arXiv:2608.16161v1cs.IRcs.AIcs.LG

TL;DR

General-purpose embeddings may confuse distinct entities with similar text and miss duplicates expressed differently. This paper fine-tunes two embedding models with domain-specific triplets, substantially improving separation between matching and challenging non-matching records.

  • Problem

    General-purpose embeddings capture semantic relatedness but provide limited evidence for distinguishing duplicate entity records from similar non-matches.

  • Method

    The study constructs identity-preserving and identity-conflicting triplets and evaluates two sentence embedding models using margin-based similarity separation.

  • Results

    Domain-specific triplet fine-tuning substantially improves duplicate-record retrieval while increasing separation from challenging non-matching candidates.

  • Takeaways & Limitations

    Targeted domain-specific supervision can reshape general-purpose embedding spaces for identity-sensitive retrieval in entity resolution.

  • Takeaways & Limitations

    The study uses synthetic records, evaluates only two embedding models, and tests controlled embedding-space separation rather than complete end-to-end entity resolution.

Abstract

from arXiv · show

General-purpose text embedding models are designed to capture semantic similarity but are not optimised for distinguishing entity records that represent the same real-world business or person. This limitation affects applications such as entity resolution and duplicate record retrieval, where small textual differences may either preserve or change identity. This paper investigates whether domain-specific triplet fine-tuning can adapt pretrained embedding models for identity-sensitive retrieval. A synthetic dataset of business and person records was created with identity-preserving variations and challenging non-matching examples. Two widely used embedding models were evaluated before and after fine-tuning using a margin-based similarity evaluation. The results show substantial improvements in separating true matches from highly similar non-matches, demonstrating that domain-specific triplet training can effectively reshape general-purpose embedding spaces for entity retrieval. These findings suggest that targeted fine-tuning provides a practical approach for improving embedding models in data quality management and information retrieval applications.

1 Introduction

Entity resolution requires embeddings that preserve identity-preserving textual variation while distinguishing subtle changes indicating different entities. This paper adapts general-purpose sentence embeddings through domain-specific triplet fine-tuning and evaluates identity-sensitive separation for business and person records.

  • Motivation: General-purpose embeddings support semantic search but are not explicitly designed to preserve identity across benign variations while separating different entities.Entity resolution therefore requires specialized representation learning that balances identity preservation with sensitivity to identity-changing textual differences.
  • Problem: Entity resolution records vary across abbreviations, spelling, formatting, incomplete addresses, punctuation, transliteration, and other source-specific differences, complicating duplicate identification.The same real-world entity may appear in multiple textual forms across operational systems, external sources, manual entry, and historical databases.
  • Approach: The paper fine-tunes two widely used sentence embedding models with domain-specific triplets built from business and person records containing identity-preserving positives and subtle identity- or location-conflicting negatives.The study evaluates models before and after fine-tuning using a margin-based protocol for identity-sensitive separation.
  • Findings: Carefully designed triplet fine-tuning substantially improves duplicate-record retrieval while increasing separation from challenging non-matching candidates.The findings show that targeted supervision can reshape general-purpose embedding spaces for entity resolution and duplicate record retrieval.

2 Dataset and experimental design

The study formulates entity resolution as similarity-ranking over business and person records, using triplets that place identity-preserving variations closer than plausible non-matches. A controlled synthetic dataset and entity-level validation split test identity-sensitive retrieval on previously unseen entities.

  • Task formulation: Entity resolution is framed as ranking records for the same real-world entity above highly similar records representing different entities.This formulation targets dense information retrieval with embedding-based candidate retrieval and ranking.
  • Learning objective: Triplet learning optimizes identity-sensitive ordering by bringing positives closer to anchors than textually similar negatives.Each triplet contains an anchor, an identity-preserving positive, and a different-entity negative; the framework follows [Schroff et al., 2015, Hoffer and Ailon, 2015].
  • Triplet design: Hard negatives share names or address components with anchors but introduce genuine identity or location conflicts, discouraging reliance on superficial textual cues.Minor punctuation, capitalization, or formatting differences were not treated as negatives without a genuine conflict.
  • Dataset construction: The synthetic dataset covers fictional business and person records with realistic names, addresses, formatting variation, and controlled identity conflicts.Synthetic generation provides control over identity and variation while avoiding confidential or personally identifiable information.
  • Data splitting and validation: Training and validation entities were separated before triplet generation, so evaluation measures generalization to previously unseen entities rather than memorization.The validation partition contained 2,000 triplets derived from entities excluded from training.

3 Methodology

The methodology fine-tuned two pretrained embedding models with triplet loss to distinguish identity-preserving variations from challenging non-matches. Performance was evaluated on 2,000 validation triplets using cosine-similarity score gaps and pass rates at three separation margins.

  • Evaluation: Evaluation used 2,000 validation triplets containing an anchor, an identity-preserving positive, and a challenging negative, with normalized embeddings compared by cosine similarity.The score gap measured separation between each positive and negative example relative to the anchor.
  • Evaluation: Pass rates were computed at score-gap margins of 0.10, 0.20, and 0.30, with larger margins representing stronger and more demanding separation.A positive score gap indicates that the matching record ranked above the non-matching record, while small gaps make retrieval and threshold selection more difficult.
  • Training: Two embedding models were independently fine-tuned with triplet loss and cosine distance to place identity-preserving records closer than challenging non-matching entities.The loss reduces anchor-positive distance while increasing anchor-negative distance when the positive is not sufficiently closer.
  • Training: Both models used the same training dataset and identical procedure, were initialized from publicly available pretrained weights, and were evaluated only on held-out validation data.Independent fine-tuning supported a fair comparison between the models.

4 Results and discussion

Triplet fine-tuning substantially improved both embedding models’ separation of identity-preserving matches from challenging non-matches across evaluation margins. BGE Base EN v1.5 gained more from adaptation and ultimately outperformed all-MiniLM-L6-v2, while the embeddings remain one component of broader deduplication workflows.

  • Pretrained performance: Pretrained models ranked identity-preserving records above challenging non-matches but provided limited separation, especially at higher score-gap margins.At margin 0.30, BGE Base EN v1.5 achieved 15.25% and all-MiniLM-L6-v2 achieved 37.85%.
  • Fine-tuned performance: At margin 0.30, fine-tuning increased BGE Base EN v1.5 from 15.25% to 92.70% and all-MiniLM-L6-v2 from 37.85% to 83.10%.Fine-tuning produced substantial improvements across all evaluated margins for both architectures.
  • Representation effects: Triplet supervision increased distances between identity-preserving and challenging non-matching records while tolerating textual variation and detecting identity or location conflicts.This clearer separation is valuable for nearest-neighbor retrieval, where small similarity differences can strongly affect candidate ranking.
  • Model comparison: After fine-tuning, BGE Base EN v1.5 outperformed all-MiniLM-L6-v2 by 9.60, 7.90, and 5.50 percentage points at margins 0.30, 0.20, and 0.10, respectively.Although all-MiniLM-L6-v2 had the stronger pretrained baseline, BGE Base EN v1.5 exhibited greater improvement after domain-specific training.
  • Operational implications: The embeddings can support duplicate retrieval, ranking, and manual-review prioritization but should be combined with identifiers, postal consistency, business rules, and review policies.Experimental margins are comparison thresholds rather than automatically selected production cut-offs.

5 Conclusion

Domain-specific triplet learning can reshape general-purpose embedding spaces to distinguish identity-preserving variations from identity or location conflicts. The study also highlights synthetic-data limitations and motivates context-conditioned embeddings for more adaptive retrieval.

  • Conclusion: Domain-specific triplet learning reshaped embedding spaces to capture application-specific similarity and distinguish identity-preserving variations from identity or location conflicts.The conclusion frames this adaptation as effective for identity-sensitive information retrieval.
  • Limitations: The study’s synthetic dataset enabled precise control over identity-preserving and identity-conflicting variations but may not represent additional noise, ambiguity, and distributional shift in enterprise data.This limitation qualifies how broadly the experimental improvements should be generalized.
  • Future directions: Embedding similarity is inherently contextual in many retrieval problems, so the same textual record may require different representations for different retrieval objectives.The conclusion challenges the assumption that a single semantic representation is sufficient across applications.
  • Future directions: Context-conditioned embedding models could combine retrieval context with textual content to generate representations guided by which input aspects should be preserved or distinguished.This direction would move beyond relying solely on fine-tuning to encode a particular similarity notion.
  • Future directions: Adaptive retrieval systems could generate multiple representations of the same text for different information needs, making efficient contextual representation learning a promising research direction.The proposed paradigm extends beyond task-specific embedding models toward more flexible information retrieval and representation learning.

Appendix A Training Configuration · Appendix B Base vs. Fine-Tuned Model Comparisons · all-MiniLM-L6-v2: Business Entities

Appendix A states that both models used the same principal fine-tuning settings, while Appendix B defines representative triplets and the reported gap for comparing pretrained and fine-tuned embeddings.

  • Appendix A Training Configuration: Both models used the same principal settings during fine-tuning.Table 5 summarizes the training configuration.
  • Appendix A Training Configuration: Table 5 provides the training configuration used in the fine-tuning procedure.
  • Appendix B Base vs. Fine-Tuned Model Comparisons: Appendix B presents representative anchor–positive–negative triplets comparing pretrained and fine-tuned embedding models.Each triplet contains an anchor record, a same-entity positive, and a near-duplicate negative distractor.
  • Appendix B Base vs. Fine-Tuned Model Comparisons: Each example pairs an anchor record with a positive representing the same entity and a negative representing a near-duplicate distractor.
  • Appendix B Base vs. Fine-Tuned Model Comparisons: The appendix defines a reported gap for comparing the positive and negative candidates.The supplied passage introduces the gap definition but does not include its formula.
  • Appendix B Base vs. Fine-Tuned Model Comparisons: A larger positive gap indicates stronger separation between the true match and the non-matching candidate.

Appendix B.1 · all-MiniLM-L6-v2: Person Entities

Appendix B.1 presents anchor–positive–negative entity examples for all-MiniLM-L6-v2, contrasting identity-preserving formatting changes with small textual differences that define negatives. The examples span names, abbreviations, punctuation, ordering, and location fields.

  • Appendix B.1: Example 1 treats “Oficina” versus “Of.” as a positive variation, while the misspelled “Cenuay” record is negative.The address remains otherwise identical across the anchor, positive, and negative records.
  • Appendix B.1: Example 2 preserves identity when address components are abbreviated and reordered, but changes the suite number to create a negative.The positive uses “Ste 288 - 334 Clover Sena Ln,” whereas the negative uses Suite 63 at a different business centre.
  • Appendix B.1: Example 3 marks abbreviations in the address as positive while treating the altered company name “Belori” as negative.The positive shortens “Suite” and “Road”; the negative retains the address but changes “Belnori.”
  • Appendix B.1: Example 4 treats punctuation, legal-form omission, and address compaction as positive, but a changed regional address as negative.The negative replaces Maldonado 20000 with Dahlia Region 10012 while retaining much of the original record.
  • Appendix B.1: Example 5 presents address reordering as positive and an inserted Paris location as negative.The positive rearranges the location components, whereas the negative adds Paris while retaining the remaining address details.

Appendix B.2 · BAAI/bge-base-en-v1.5: Business Entities

Appendix B.2 presents business-entity triplets for BAAI/bge-base-en-v1.5, pairing identity-preserving record variations with non-matching records sharing the same person name. The examples vary abbreviated names, reordered addresses, and changed locations.

  • Appendix B.2: The appendix uses anchor-positive-negative triplets in which positives preserve identity while negatives retain similar names but change the associated address.This structure is explicit in the Goran Valente, Nuru Tadesse, Parisa Dervishi, and Leire Okafor examples.
  • Appendix B.2: Positive examples preserve identity through abbreviated names, including “Goran V.”, “Nuru T.”, “Parisa D.”, and “Leire O.”.Each abbreviation is paired with the same stated residence as its anchor, apart from omitted postal detail in several examples.
  • Appendix B.2: Negative examples keep the person name but assign a different full address, as shown for Goran Valente, Nuru Tadesse, Parisa Dervishi, and Leire Okafor.The negative locations differ in unit, street, city, region, or country from the corresponding anchors.
  • Appendix B.2: The examples also test address-format variation: Uche Eriksen’s positive reorders the address fields while preserving the same location and identity.The corresponding negative keeps the same name and unit but changes the city and region.

Appendix B.3 · BAAI/bge-base-en-v1.5: Person Entities

Appendix B.3 presents person-entity evaluation examples for BAAI/bge-base-en-v1.5 using anchors, identity-preserving positives, and challenging negatives that alter location details. The examples span firms across multiple countries and include spelling, abbreviation, punctuation, and address variations.

  • Appendix B.3: The evaluation pairs each anchor with a lightly altered positive and a highly similar negative whose location changes to another city, region, or country.This structure tests whether identity-preserving textual variation is distinguished from geographically inconsistent records.
  • Appendix B.3: Example 1 preserves the Estonian firm identity despite spelling and address abbreviations, while the negative retains the name but changes the address to Al Ain, United Arab Emirates.The positive changes “Socety” and shortens “tanav,” whereas the negative replaces the location with Al Ain and a different country.
  • Appendix B.3: Across the examples, the reported evaluation format is organized by stage with positive similarity, negative similarity, and gap fields.The supplied passages show the metric headings but do not provide numerical values for these fields.
  • Appendix B.3: Example 2 preserves the Bristol firm through spelling, suite, street, and postcode variations, while the negative changes the location to Osaka, Japan.The positive uses “Socety,” “Ste,” “Ln,” and a shortened postcode; the negative keeps the business name but changes the address country.
  • Appendix B.3: Example 3 preserves the Costa Rican firm despite abbreviating the legal form and address, while the negative changes the location to Bedok, Singapore.The positive shortens “SR.L.” and address terms; the negative retains the firm identity but substitutes a Singaporean location.
  • Appendix B.3: Example 4 preserves the Polish firm despite removing punctuation and abbreviating the address, while the negative changes the location to Portland, United States.The positive alters the company suffix and street notation; the negative changes the city and country while retaining the business name.
  • Appendix B.3: Example 5 applies the same pattern to an Uzbek firm, pairing abbreviated spelling and address details with a negative that changes the location to Munich, Germany.This lower-salience example preserves Bukhara and Uzbekistan in the positive while replacing them with Munich and Germany in the negative.

Appendix B.4 … Appendix B.6

The appendices illustrate identity-preserving address variations and challenging non-matches across representative person records. Table 6 summarizes the average discrimination gap, defined as the difference between fine-tuned and pretrained average gaps.

  • Appendix B.4: Address-order changes form positive matches, while changed cities or regions form negative examples for Gita Gyamfi and Zoran Deshmukh.The examples preserve the person and address content in positives but alter location information in negatives.
  • Appendix B.4: Abbreviating “Lot” to “Lt” preserves the positive identity example, whereas replacing the address with a different unit and street creates a negative.This example retains the same city, region, postal code, and person name while varying the address substantially in the negative.
  • Appendix B.4: An exact-match positive for Tenzin Etxeberria contrasts with a negative that keeps the name and apartment but changes the city, region, postal code, and country.The positive is explicitly labeled “exact match.”
  • Appendix B.4: Mirela Haddad’s example similarly treats reordered address components as positive and different city and region values as negative.This example is included among the representative cases but is marked with scrim salience.
  • Average Discrimination Gap: Table 6 summarizes the average discrimination gap across representative business and person examples, comparing fine-tuned and pretrained average gaps.The improvement is calculated as the difference between the fine-tuned and pretrained average gaps.

Observations

Fine-tuning consistently widened the positive–negative gap across the evaluated models and entity-type groups, with the largest improvement for BAAI/bge-base-en-v1.5 on business entities.

  • Observations: Fine-tuning widens the positive–negative gap for every model and entity-type group shown, sharpening discrimination between true matches and near-duplicate distractors.The comparisons reveal a consistent improvement across the evaluated groups.
  • Observations: The BAAI/bge-base-en-v1.5 business-entity gap increases from approximately 0.069 before fine-tuning to 1.287 afterward, the largest improvement observed.This comparison is identified as the strongest effect among the representative results.
  • Observations: Fine-tuning also produces a substantial improvement for all-MiniLM-L6-v2 across the evaluated entity-resolution comparisons.The supplied passage states that the improvement applies to all-MiniLM-L6-v2, but does not provide its numerical values here.
Loading 2608.16161v1…