Generative Language Model Bias Mapping for Multi-Party Summaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing generative language models struggle with accurately summarizing multi-party communications due to domain-specific biases in their training data, leading to inefficiencies and reduced accuracy when applied to domains other than their training data.
Innovation Solution
A computerized method using a generative language model (GLM) that incorporates a defined set of training bias words to control and reduce the generation of domain-specific words, guiding the model to generate alternative summarization words by mapping them to a curated list of training bias words, thereby enhancing accuracy and flexibility across different domains.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a generative language model is trained with domain-specific training data, then it can generate accurate summaries for that specific domain, but it develops biases that reduce its accuracy and flexibility when applied to other domains
Solution Approach 1:
The patent extracts and removes domain-specific biased words from the model's output by mapping generated words against a curated list of training bias words. When a biased word is detected, it is replaced with an alternative word that is less domain-specific, thereby removing the harmful bias while preserving the summarization function.
Solution Approach 2:
The patent changes the parameter of word selection by introducing a bias-reduction mechanism that modifies the probability distribution of word choices. The model is guided to select alternative words that have lower domain-specificity scores, effectively changing the parameter of domain specificity in the generated output.
2Device complexity
If existing text summarization methods are used for multi-party communications, then processing is simpler, but accuracy is significantly reduced due to inability to handle multi-party dynamics
Solution Approach 1:
The patent segments multi-party communications into distinct speaker turns and identifies party identities for each segment. This segmentation allows the model to track which party is speaking and maintain accurate attribution throughout the summary generation process, handling multi-party dynamics systematically.
Solution Approach 2:
The patent introduces an intermediary bias-reduction layer that sits between the base summarization model and the final output. This intermediary component maps generated words against training bias lists and performs word substitution, acting as a mediator that cleanses the output without requiring fundamental changes to the underlying summarization architecture.
3Measurement precision
If domain-specific training data is used extensively, then model accuracy for that domain improves, but time and resources required for training and adaptation increase significantly
Solution Approach 1:
The patent creates a universal bias-reduction mechanism that can be applied across multiple domains without requiring domain-specific retraining. The curated list of training bias words serves as a general resource that reduces domain-specific biases in any domain, making the model universally applicable while maintaining accuracy.
Solution Approach 2:
The patent performs preliminary action by curating lists of domain-specific biased words before the actual summarization task. This pre-processing of bias identification allows the model to efficiently reduce biases during inference without requiring extensive domain-specific training, saving time and computational resources.
Data Source
AI summary
The disclosure herein describes reducing training bias in outputs generated by a generative language model. A communication segment associated with a communication is obtained by at least one processor of a generative language model. An output value associated with the communication segment is generated by the generative language model. The output value is mapped to a set of training bias values associated with the generative language model and based on the mapping of the output value to a training bias value of the set of training bias values, an alternative output value is generated. The alternative output value is used in a generated segment output for the communication segment. The accuracy of segment outputs generated by the generative language model is improved through reducing or eliminating its training biases.