Conditional Language Model Training via Classifier-Guided Data Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computer-implemented conditional language models (CLMs) face suboptimal performance due to spurious correlations in training data, particularly when data is unbalanced, leading to poor controllability in generating output texts based on specified attributes.
Innovation Solution
A method to generate balanced training data by using a first CLM to produce output texts under various conditions, with a classifier identifying attribute values, and only including matching outputs in the training data for a second CLM, thereby reducing spurious correlations and improving controllability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional training data is used to train CLM, then the model can be trained with existing data, but the training data is unbalanced causing spurious correlations and suboptimal performance
Solution Approach 1:
The patent applies preliminary action by using a first CLM to pre-generate balanced training data before training the second CLM. The first CLM generates output texts for various conditions, a classifier validates attribute matching, and only matched outputs are included in the training data, thereby preparing clean, balanced data in advance to prevent spurious correlations from forming during the second CLM's training.
Solution Approach 2:
The patent implements feedback through a classifier that validates the attribute values of output texts generated by the first CLM. The classifier compares actual attribute values with specified conditions and provides feedback by filtering out mismatched examples. This feedback mechanism ensures only high-quality, attribute-matched data is used for training the second CLM, eliminating spurious correlations.
2Adaptability or versatility
If unbalanced training data is used, then more diverse data can be included, but categories with fewer samples perform sub-optimally
Solution Approach 1:
The patent applies local quality by generating and validating training data specifically for each condition category. The first CLM generates output texts for multiple conditions, the classifier validates attribute matching for each condition individually, and the training data is constructed to ensure balanced representation across all categories. This localized validation ensures each category has sufficient quality samples for optimal performance.
3Measurement precision
If the training data generation process includes classification and filtering, then data quality improves, but the process complexity increases
Solution Approach 1:
The patent uses an intermediary approach by introducing a classifier as a mediator between the first CLM and the training data construction process. The classifier validates attribute values and filters output texts, acting as an intermediate validation layer. This intermediary component simplifies the overall process by providing a clear decision boundary for data inclusion, making the complex data generation process more manageable and precise.
Data Source
AI summary
Technologies related to computer-implemented conditional language models (CLMs) are described. A first CLM is trained to generate output texts based upon input texts and conditions. Output texts generated by the first CLM are included in a training set, and a second CLM is trained based upon the training set. The second CLM is then configured to receive input text and a condition and generate an output text based upon the input text and the condition.