Negative Language Model for Evaluating Grammaticality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current language models are susceptible to generating low perplexity scores for ungrammatical text due to the presence of negative signals, which leads to false positives and inability to distinguish between grammatical and ungrammatical inputs, as they are primarily trained on positive text data without effective evaluation for undesirable outputs.
Innovation Solution
A method involving a 'negative' language model is introduced, which generates ungrammatical outputs and is trained on a separate corpus, allowing for the evaluation of target language models by testing their performance with negative data to detect and mitigate undesirable signals, using techniques such as perplexity scoring and comparison with thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If language models are trained on large corpora of positive text data to maximize prediction accuracy for grammatical text, then their ability to predict positive text is improved, but they become susceptible to generating low perplexity scores for ungrammatical text due to negative signals
Solution Approach 1:
The patent applies preliminary anti-action by training a negative language model on corrupted text data before using it to generate negative test data. This negative model learns to produce ungrammatical outputs that contain negative signals, which are then used to evaluate and improve the target language model's robustness against such signals, preventing false positive predictions
Solution Approach 2:
The negative language model serves as an intermediary that bridges the gap between positive training data and the need to evaluate negative signal susceptibility. It generates intermediate negative test data that allows the target model to be assessed and improved for reliability without requiring manually crafted ungrammatical examples
2Quantity of substance
If language models are trained only on positive text data, then training data availability is improved, but evaluation capability for undesirable outputs deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-training the negative language model on corrupted versions of the positive training corpus before using it to generate negative test data. This preliminary corruption process creates a reusable negative data generator that enables ongoing evaluation of the target model's susceptibility to negative signals without requiring manual intervention
Solution Approach 2:
The negative language model is trained using automated text corruption techniques applied to the existing positive corpus, allowing the system to self-generate negative test data without external manual input. The model serves itself by converting available positive data into negative evaluation cases through programmatic corruption methods
Data Source
AI summary
A method of evaluating a language model using negative data may include accessing a first language model that is trained using a first training corpus, and accessing a second language model. The second language model may be configured to generate outputs that are less grammatical than outputs generated by the first language model. The method may also include training the second language model using a second training corpus, and generating output text from the second language model. The method may further include testing the first language model using the output text from the second language model.