Machine Learning Data Classification via Segmented Prompting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face challenges in performing Chain-of-Thought (CoT) responses effectively, particularly when requiring both primary prediction and justification, due to increased cognitive load and error propagation in single-prompt schemes.
Innovation Solution
The method involves dividing the data classification task into two separate tasks: one for classification and another for generating Chain-of-Thought reasoning. This is achieved by creating two distinct samples for training, each focused on either the classification or the reasoning aspect, allowing the machine learning model to optimize each task independently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single prompt is used for both classification and Chain-of-Thought reasoning, then the model can perform both tasks in one pass, but the cognitive load increases and error propagation occurs
Solution Approach 1:
The patent divides the single prompt task into two separate prompts: the first prompt handles classification prediction, and the second prompt handles Chain-of-Thought reasoning based on the classification result. This segmentation reduces cognitive load on the model and prevents error propagation between tasks, thereby improving reasoning accuracy while maintaining task completion efficiency.
2Adaptability or versatility
If the model is trained to provide both classification and reasoning in one response, then the output is comprehensive, but the precision of both classification and reasoning decreases
Solution Approach 1:
The patent separates the classification and reasoning tasks into two distinct prompts, allowing each to be optimized independently. The first prompt focuses solely on classification precision, while the second prompt focuses on reasoning quality. This segmentation maintains response completeness while improving the precision of both individual tasks compared to attempting both in a single response.
3Loss of information
If Chain-of-Thought reasoning is required along with primary prediction, then the model provides more explanatory power, but the cognitive burden on the model increases
Solution Approach 1:
The patent divides the cognitive task into two separate prompts: the first prompt handles the primary classification prediction with lower cognitive burden, and the second prompt handles the Chain-of-Thought reasoning based on the classification result. This segmentation reduces the cognitive complexity required for each individual prompt while maintaining complete information output through the combined responses.
Solution Approach 2:
The first prompt performs the classification prediction as a preliminary action before the second prompt generates the Chain-of-Thought reasoning. This preliminary classification provides a foundation that simplifies the subsequent reasoning task, reducing the overall cognitive burden on the model while maintaining information completeness.
Data Source
AI summary
There are provided methods, devices, and computer program products for data classification. A sample for training a machine learning model is obtained. The sample comprises a prompt and a response for the prompt, the prompt comprises input data, and the response comprises a classification of the input data, and a reason why the input data belongs to the classification. A first sample is determined based on the input data and the classification of the input data, and the first sample comprises a first prompt and a first response. A second sample is determined based on the input data, the classification of the input data, and the reason, and the second sample comprises a second prompt and a second response. The machine learning model is updated based on the first and the second samples. Therefore, the machine learning model may be updated in a more reliable and accurate way.