Adversarial Token Detection Network for NLP Training Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning methods using deep learning in natural language processing face challenges in achieving high accuracy due to the generator's focus on masked characters, leading to decreased learning efficiency in the latter half of the process, as the correct answer rate increases and the discriminator learns to identify original answers more frequently, resulting in a deterioration of learning efficiency.
Innovation Solution
The proposed solution involves an adversarial replaced token detection network with a generator and discriminator, where pre-learning uses unsupervised training data without correct answer information, and fine-tuning employs supervised training data with correct answers, generating error information to improve the accuracy of the machine learning model by constructing a topology that challenges the discriminator, and incorporating a restorer to maintain sentence consistency in the second embodiment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If the generator focuses on masked characters during training, then the training process can be simplified, but the learning efficiency decreases in the latter half of the process as the discriminator learns to identify original answers more frequently
Solution Approach 1:
The patent introduces a feedback mechanism where the discriminator's output is fed back to the generator. The generator receives not only masked characters but also feedback information from the discriminator about whether the generated tokens are original or replaced. This feedback loop allows the generator to adjust its training strategy dynamically, preventing the efficiency degradation that occurs when the discriminator becomes too good at identifying original answers.
Solution Approach 2:
The patent changes the training parameters by introducing a mixing ratio parameter that controls the proportion of different training samples. The system dynamically adjusts this parameter based on training progress, switching between different training modes (character-level and word-level) to maintain optimal learning efficiency throughout the training process.
2Measurement precision
If pre-learning uses a huge amount of data to achieve high accuracy, then the model accuracy improves, but the arithmetic amount becomes huge and processing time increases
Solution Approach 1:
The patent segments the training process into two distinct stages: pre-learning and fine-tuning. During pre-learning, the model is trained on a large corpus of text data to learn general language patterns and representations. This preliminary training on diverse data establishes a strong foundation. Subsequently, the model undergoes fine-tuning on a smaller, task-specific dataset where the actual target problem is solved. This segmentation allows the model to achieve high accuracy without requiring all training data to be processed simultaneously in the final stage.
Solution Approach 2:
The pre-learning stage performs preliminary action by learning general language representations and patterns from large amounts of data before the actual fine-tuning task. This preliminary learning creates efficient representations that can be quickly adapted to specific tasks during fine-tuning, reducing the amount of data and computation needed in the final training stage while maintaining high accuracy.
3Ease of operation
If the generator generates second input data by rewriting parts of first input data, then the training can proceed with unsupervised learning, but the discriminator may learn to identify original answers more frequently, reducing learning efficiency
Solution Approach 1:
The patent introduces asymmetry in the training process by using different loss functions and training objectives for the generator and discriminator. The generator is trained with a loss function that penalizes both incorrect replacements and failure to replace when needed, while the discriminator uses a loss function that balances the detection of replaced tokens with the preservation of original ones. This asymmetric training approach prevents the discriminator from becoming too efficient at identifying original answers.
Solution Approach 2:
The training process is made dynamic by introducing adaptive learning rates and dynamic mixing ratios that change during training. The system automatically adjusts these parameters based on training progress and performance metrics, allowing the training to adapt to changing conditions and maintain optimal learning efficiency throughout the process.
Data Source
AI summary
A non-transitory computer-readable recording medium stores a program for causing a computer to execute a process, the process includes inputting training data to a machine learning model that includes a generator and a discriminator, the generator generating second input data in which a part of first input data is rewritten in response to an input of the first input data, the discriminator discriminating a rewritten portion in response to an input of the second input data generated by the generator, generating correct answer information, based on the training data and an output result of the generator, and executing training of the machine learning model by using first error information obtained based on the output result of the generator and a discrimination result of the discriminator, and second error information obtained based on the discrimination result of the discriminator and the correct answer information.