Adaptive NER Training Data Balancing for Minority Entity Classes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional named entity recognition (NER) models are biased towards learning majority class entities due to unbalanced representations in training datasets, leading to misclassification of minority class entities, and existing augmentation techniques further skew the datasets, impacting performance on minority classes.
Innovation Solution
Adaptive training data augmentation techniques calculate separate augmentation rates for each named entity category based on their distribution in the training data, generating adaptive numbers of additional utterances to normalize entity distributions and reduce bias, thereby improving NER model performance on minority classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional NER models are trained on unbalanced training datasets, then the model learns majority class entities well, but misclassifies minority class entities
Solution Approach 1:
The patent applies local quality by calculating and applying different augmentation rates for different named entity categories based on their distribution in the training data. Minority classes receive higher augmentation rates while majority classes receive lower rates, creating localized adjustments to the training data distribution that address the specific imbalance problem without affecting all classes uniformly.
Solution Approach 2:
The patent changes the parameter of data distribution by dynamically adjusting augmentation rates for different entity categories. The system modifies the training data composition by generating additional samples at different rates for different classes, thereby transforming the unbalanced distribution into a more balanced one that improves minority class recognition while maintaining overall classification accuracy.
2Quantity of substance
If existing data augmentation techniques are applied to NER training data, then the quantity of training data increases, but the dataset becomes further skewed, impacting performance on minority classes
Solution Approach 1:
The patent applies dynamics by making the augmentation rate adaptive and dynamic rather than static. The system calculates augmentation rates based on the actual distribution of entity categories in the training data, allowing the augmentation strategy to dynamically adjust to the specific imbalance characteristics of each dataset. This dynamic approach ensures that augmentation benefits minority classes without creating new imbalances.
Solution Approach 2:
The patent changes the parameter of data distribution by applying different augmentation rates to different entity categories. Instead of uniform augmentation that would further skew the dataset, the system transforms the distribution by generating more samples for minority classes and fewer or no additional samples for majority classes, thereby improving class balance while increasing overall data volume.
3Quantity of substance
If uniform data augmentation is applied to all entity categories, then the training data volume increases uniformly, but the bias towards majority classes is exacerbated
Solution Approach 1:
The patent applies local quality by implementing category-specific augmentation rates rather than uniform augmentation. The system identifies minority classes that need more training samples and applies higher augmentation rates to them, while applying lower or zero augmentation rates to majority classes. This localized approach ensures that training data volume increases where it is most needed without exacerbating existing biases.
Solution Approach 2:
The patent changes the parameter of augmentation intensity by applying different rates to different entity categories. The system transforms the uniform augmentation parameter into a differentiated parameter set where each category receives augmentation proportional to its scarcity, thereby increasing overall data volume while simultaneously improving minority class recognition accuracy rather than worsening it.
Data Source
AI summary
Techniques are disclosed herein for adaptive training data augmentation to facilitate training named entity recognition (NER) models. Adaptive augmentation techniques are disclosed herein that take into consideration the distribution of different entity types within training data. The adaptive augmentation techniques generate adaptive numbers of augmented examples (e.g., utterances) based on the distribution of entities to make sure enough numbers of examples for minority class entities are generated during augmentation of the training data.