Transformer Entity Matching With Joint NER Loss for Record Linking

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Cloud platforms face significant overhead in translating and connecting data for entities across different applications, data models, and databases due to disconnected storage of information about the same entity, leading to redundancies and inefficiencies.

Innovation Solution

A machine-learned entity matching model, configured as a transformer architecture, is trained using a combination of a first loss for entity matching predictions and a second loss for named-entity recognition (NER) predictions to identify and distinguish between entities, reducing redundancies by merging records that refer to the same entity.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If data about the same entity is stored across different applications, data models, and databases in a disconnected manner, then data can be stored using different data models for different business units and tenants, but significant overhead is spent on translating and connecting data for entities

Engineering Contradiction:

Improveability to store data using different data modelsVSAvoidoverhead for translating and connecting data

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent introduces an entity matching model as an intermediary component that automatically identifies and links records referring to the same entity across different applications and data models. This mediator translates disconnected data representations into connected entity relationships, eliminating the manual overhead of data translation and connection while preserving the ability to use different data models for different business units and tenants

Inventive Principle:

Principle #24Intermediary (Mediator)

2Measurement precision

If entity matching accuracy is improved by training with multiple loss functions, then the model can distinguish between similar and different entities, but the training complexity and computational resources increase

Engineering Contradiction:

Improveentity matching accuracyVSAvoidtraining model complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges multiple loss functions (entity matching loss and named entity recognition loss) into a unified training objective. This combination allows the model to simultaneously learn entity matching accuracy and entity type recognition, improving overall matching precision while managing training complexity through a integrated loss function framework rather than separate complex training processes

Inventive Principle:

Principle #5Merging (Combining)

Data Source

PatentUS12608549B2Training machine learning models for multi-modal entity matching in electronic records

Publication Date: 2026.04.21 SALESFORCE INC

AI summary

A cloud platform trains a machine-learned entity matching model that generates predictions on whether a pair of electronic records refer to a same entity. In one embodiment, the entity matching model is configured as a transformer architecture. In one instance, the entity matching model is trained using a combination of a first loss and a second loss. The first loss indicates a difference between an entity matching prediction for a training instance and a respective match label for the training instance. The second loss indicates a difference between a set of named-entity recognition (NER) predictions for the training instance and the set of NER labels for the tokens of the training instance.