Handwriting Recognition With Selective Language Model Refinement

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing handwriting recognition systems struggle with historical documents due to degraded records, language changes, varied handwriting styles, and spelling errors, leading to inaccurate transcriptions and high labor costs.

Innovation Solution

Incorporating a character-based language model into handwriting recognition systems to refine predictions by leveraging knowledge of spoken and written language, using a loss prediction module to determine when to apply the language model for improved accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a language model is integrated into the handwriting recognition system to improve accuracy, then the robustness and word accuracy are improved, but the device complexity and computational resources increase

Engineering Contradiction:
ImproverobustnessVSAvoiddevice complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system performs preliminary loss prediction using a lightweight module before deciding whether to invoke the full language model. This preliminary action filters out cases that don't require complex processing, thereby reducing overall system complexity while maintaining robustness where needed.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The handwriting recognition system is segmented into distinct components: a base recognition model, a loss prediction module, and a language model. This segmentation allows selective application of computational resources, where the language model is only activated when the loss prediction indicates potential errors, thus managing device complexity effectively.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If a language model is used to refine handwriting predictions, then the word accuracy is improved, but the computational resources and processing time increase

Engineering Contradiction:
Improveword accuracyVSAvoidcomputational resources
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

Instead of applying the language model to all predictions, the system uses partial action by selectively applying it only to cases where the loss prediction module indicates potential errors. This reduces computational resource consumption while maintaining high word accuracy where it matters most.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The loss prediction module performs a preliminary assessment before the language model is invoked. This preliminary filtering action identifies only those predictions that benefit from language model refinement, thereby optimizing computational resource usage while preserving word accuracy.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If the system processes all handwriting predictions through the language model, then the word accuracy is maximized, but the productivity and processing speed decrease

Engineering Contradiction:
Improveword accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The system applies partial action by processing only a subset of predictions through the language model - specifically those flagged by the loss prediction module. This maintains high word accuracy for problematic cases while preserving processing speed for the majority of straightforward predictions.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The processing pipeline is segmented into fast path (direct output for high-confidence predictions) and slow path (language model refinement for uncertain predictions). This segmentation enables the system to maintain high productivity for most cases while achieving maximum word accuracy when needed.

Inventive Principle:
Principle #1Segmentation

4Reliability

If the language model is applied to correct spelling errors and unrecognized words, then the robustness is improved, but the device complexity increases

Engineering Contradiction:
ImproverobustnessVSAvoiddevice complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The loss prediction module performs a preliminary screening to identify predictions that may contain spelling errors or unrecognized words. This preliminary action enables the robustness improvement from the language model to be applied only where necessary, thereby limiting the increase in device complexity.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The loss prediction module acts as an intermediary between the base recognition model and the language model. It mediates by filtering and directing only appropriate cases to the language model, thus enabling robustness improvement while controlling device complexity through selective intermediation.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4238068B1Improving handwriting recognition with language modeling
Publication Date: 2025.09.10 ANCESTRY COM OPERATIONS INC
  • EP4238068B1 patent drawingFigure 1
  • EP4238068B1 patent drawingFigure 2
  • EP4238068B1 patent drawingFigure 3

AI summary

Systems and methods for handwriting recognition using language modeling facilitate improved results by using a trained language model (276) to improve results from a handwriting recognition machine learning model (204). The language model (276) may be a character-based language model trained on a dataset pertinent to field values on which the handwriting recognition model (204) is to be used. A loss prediction module (256) may be trained with the handwriting recognition model (204) and/or the language model (276) and used to determine whether a prediction (210) from the handwriting recognition model (204) should be refined by passing the prediction (210) through the trained language model (276).