Self-Cleaning Discriminator for Noisy NER Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional named entity recognition (NER) models face inefficiencies in training due to the use of noisy training data, leading to inaccurate and computationally expensive training processes, and are often rigid in adapting to specific tasks without access to high-quality training data.
Innovation Solution
A self-cleaning guided denoising framework using a discriminator model to correct noise in noisy training data by reweighting labels, enabling flexible and efficient training of NER models on large, noisy datasets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional NER models use noisy training data, then the quantity of training data increases, but the accuracy of the NER model deteriorates
Solution Approach 1:
The patent introduces a self-training discriminator model as an intermediary component that mediates between the noisy training data and the NER model. The discriminator model processes the noisy data first, generating cleaned training samples with corrected labels, which then serve as input to the NER model. This intermediary processing step enables the system to utilize large quantities of noisy training data while maintaining high accuracy by filtering out noise before the main learning process.
2Quantity of substance
If conventional NER models use noisy training data, then more training data can be processed, but the training process becomes computationally expensive
Solution Approach 1:
The patent applies preliminary action by implementing a self-training phase where the discriminator model is first trained on a small subset of clean data to learn noise patterns. This preliminary training enables the discriminator to efficiently identify and correct noise in larger datasets without requiring computationally intensive processing during the main NER training phase. The pre-learned noise detection capabilities significantly reduce the computational burden of processing large noisy datasets.
3Manufacturing precision
If conventional NER models are trained with high-quality data, then accuracy improves, but adaptability to specific tasks deteriorates
Solution Approach 1:
The patent implements universality by designing a discriminator model with multi-functional capabilities. The same discriminator model can process noisy data from different domains and task types while maintaining its noise detection and correction functionality. This universal approach allows the NER system to adapt to specific tasks by simply changing the training data source rather than requiring task-specific model architectures, thereby achieving both high accuracy and versatility across different NER applications.
Data Source
AI summary
This disclosure describes one or more implementations of systems, non-transitory computer-readable media, and methods that train a named entity recognition (NER) model with noisy training data through a self-cleaning discriminator model. For example, the disclosed systems utilize a self-cleaning guided denoising framework to improve NER learning on noisy training data via a guidance training set. In one or more implementations, the disclosed systems utilize, within the denoising framework, an auxiliary discriminator model to correct noise in the noisy training data while training an NER model through the noisy training data. For example, while training the NER model to predict labels from the noisy training data, the disclosed systems utilize a discriminator model to detect noisy NER labels and reweight the noisy NER labels provided for training in the NER model.