Focused Language Model Training with Unlabeled and Labeled Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence-based chatbots are difficult and costly to develop due to the need for specialized knowledge and techniques, and publicly available pre-trained language models are not suitable for all natural language processing tasks, requiring custom models that are inefficient to generate.
Innovation Solution
A framework for focused training of language models involves post-training a pre-trained model with unlabeled and labeled data to optimize model parameters for a target domain, task, or language, and using iterative training operations with multi-task learning to enhance performance on downstream tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a custom natural language processing model is generated from scratch, then the model can be optimized for specific tasks and domains, but the development time, cost, and computational resources required are extremely high
Solution Approach 1:
The patent applies preliminary action by using pre-trained language models that have already undergone extensive training on large corpora before being made available for use. This pre-training phase performs the computationally intensive work in advance, allowing users to leverage these models for specific tasks without repeating the entire training process. The framework enables task-specific adaptation through fine-tuning or prompt engineering on top of these pre-trained foundations, dramatically reducing development time while maintaining task optimization capabilities.
Solution Approach 2:
The patent employs copying by creating task-specific variants or adaptations of general-purpose pre-trained language models. Instead of building models from scratch, the approach copies the knowledge and capabilities already embedded in pre-trained models and adapts them to specific domains and tasks through targeted fine-tuning, parameter adjustment, or prompt design. This allows multiple task-specific models to be derived from a single pre-trained foundation.
2Manufacturing precision
If a custom natural language processing model is generated from scratch, then the model can be optimized for specific tasks and domains, but the computational resources and expense required are extremely high
Solution Approach 1:
The patent applies preliminary action by using pre-trained language models that have already undergone extensive training on large corpora before being made available for use. This pre-training phase performs the computationally intensive work in advance, allowing users to leverage these models for specific tasks without repeating the entire training process. The framework enables task-specific adaptation through fine-tuning or prompt engineering on top of these pre-trained foundations, dramatically reducing development time while maintaining task optimization capabilities.
Solution Approach 2:
The patent employs copying by creating task-specific variants or adaptations of general-purpose pre-trained language models. Instead of building models from scratch, the approach copies the knowledge and capabilities already embedded in pre-trained models and adapts them to specific domains and tasks through targeted fine-tuning, parameter adjustment, or prompt design. This allows multiple task-specific models to be derived from a single pre-trained foundation.
3Productivity
If publicly available pre-trained language models are used, then development efficiency is improved, but the models are not suitable for all natural language processing tasks and require customization
Solution Approach 1:
The patent applies dynamics by creating a flexible framework that allows pre-trained language models to be dynamically adapted to different tasks and domains. The system enables users to adjust model parameters, fine-tune on task-specific data, modify prompts, and combine multiple models as needed. This dynamic adaptation capability allows a single pre-trained model foundation to serve multiple different NLP tasks effectively, bridging the gap between general-purpose pre-training and task-specific requirements.
Solution Approach 2:
The patent employs universality by designing a framework where pre-trained language models serve as multi-functional foundations that can be adapted to various NLP tasks. The system enables a single pre-trained model to perform multiple functions through techniques like prompt engineering, parameter adjustment, and fine-tuning on different task datasets. This multi-functionality approach allows one model to replace what would traditionally require multiple specialized models.
Data Source
AI summary
Techniques are disclosed herein for focused training of language models and end-to-end hypertuning of the framework. In one aspect, a method is provided that includes obtaining a machine learning model pre-trained for language modeling, and post-training the machine learning model for various tasks to generate a focused machine learning model. The post-training includes: (i) training the machine learning model on an unlabeled set of training data pertaining to a task that the machine learning model was pre-trained for as part of the language modeling, and the unlabeled set of training data is obtained with respect to a target domain, a target task, or a target language, and (ii) training the machine learning model on a labeled set of training data that pertains to another task that is an auxiliary task related to a downstream task to be performed using the machine learning model or output from the machine learning model.