# Beyond Symmetry: Asymmetric Matching with Transformers for Collective Entity Resolution

DevFeed: [Beyond Symmetry: Asymmetric Matching with Transformers for Collective Entity Resolution](<https://devfeed.tech/articles/beyond-symmetry-asymmetric-matching-with-transformers-for-collective-entity-resolution-132692.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.15185>)

Author: Andrei Olar

Published: 2026-11-13T05:00:00Z

Content type: opinion

Language: en

Sources: [arXiv: Databases](<https://devfeed.tech/sources/arxiv-databases.md>)

Topics: [Entity resolution](<https://devfeed.tech/topics/entity-resolution.md>), [Transformers](<https://devfeed.tech/topics/transformers.md>)

Tags: [collective](<https://devfeed.tech/tags/collective.md>), [cs-db](<https://devfeed.tech/tags/cs-db.md>), [entity-resolution](<https://devfeed.tech/tags/entity-resolution.md>), [matching](<https://devfeed.tech/tags/matching.md>), [symmetry](<https://devfeed.tech/tags/symmetry.md>), [transformers](<https://devfeed.tech/tags/transformers.md>)

## AI overview

The paper finds that input order can affect entity resolution matching and skew the final clusters. It studies a collective framework with separate matching and clustering stages, using pretrained transformers in an order-aware matcher to model asymmetry and measure its effects across benchmark datasets. The abstract reports that asymmetry is measurable and argues supervised matchers may need direction-aware training that preserves it, especially when information is vague or ambiguous. The supplied evidence does not give benchmark details or quantify the reported skew.

## Source excerpt

arXiv:2611.15185v1 Announce Type: new Abstract: A core assumption of entity resolution (ER) is its symmetric nature, which rests on a principle so self-evident it borders on tautology: if a refers to the same object as b, then b must also refer to the same object as a. However, this global property of symmetry does not necessarily extend to each of the complex decision processes within ER systems. In cases involving vague or ambiguous information, enforcing symmetry can be detrimental to performance, an issue exemplified by the ``bad triplet problem''. This paper investigates the effects of input-order asymmetry within a collective ER framework that integrates distinct matching and clustering stages. The input order-aware matching stage utilizes pre-trained transformers to explicitly model and exploit asymmetry, making it possible to evaluate its effect on clustering performance across multiple benchmark datasets. Our results demonstrate that asymmetry is not only measurable but also introduces a significant skew in the final entity clusters, highlighting the need to re-evaluate supervised training methodologies for matchers, particularly the development of direction-aware training paradigms that can preserve, rather than neutralize, the inherent asymmetry in matching tasks.