Masked Language Model Screening for NLP Noise Samples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of obtaining high-quality training samples is a significant obstacle in natural language processing applications, particularly for deep learning models, as it affects the performance and implementation of these models in practical applications.
Innovation Solution
A method is proposed to identify noise samples by obtaining an original sample set, creating a target sample set through a preset adjustment rule, performing mask prediction using a pre-trained language model, and matching the prediction results with a preset condition to determine noise samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manual inspection of training samples is used, then quality control can be performed, but human resources are consumed and efficiency is low
Solution Approach 1:
The system uses a pre-trained language model to automatically inspect and identify noise samples in the training set. The model performs mask prediction on training corpora and compares predictions with actual labels to automatically flag noise samples, eliminating the need for manual human inspection while maintaining quality control capabilities.
Solution Approach 2:
The patent replaces manual human inspection with an automated computational system. The language model processes training samples through mask prediction mechanisms, automatically detecting inconsistencies between predicted and actual labels, thereby substituting mechanical human labor with automated digital processing.
2Reliability
If all training samples are used, then dataset size is maximized, but noise samples reduce model performance
Solution Approach 1:
The system extracts and identifies noise samples from the training set by comparing mask prediction results with actual labels. Once identified, these noise samples can be removed or corrected, separating high-quality useful samples from harmful noise to improve overall training data quality while maintaining dataset size.
Solution Approach 2:
The patent implements a feedback mechanism where the language model's mask prediction results are compared with actual training labels. This feedback loop identifies inconsistencies that indicate noise samples, allowing the system to iteratively improve training data quality by flagging and removing problematic samples.
Data Source
AI summary
The method for identifying noise samples, includes: obtaining an original sample set; obtaining a target sample set by adding masks to original training corpora in the original sample set using a preset adjustment rule; performing mask prediction on a plurality of target training corpora in the target sample set using a pre-trained language model to obtain a first mask prediction character corresponding to each target training corpus; matching the first mask prediction character corresponding to each target training corpus with a preset condition; and according to target training corpora of which first mask prediction characters do not match the preset condition in the target sample set, determining corresponding original training corpora in the original sample set as noise samples.