Generative Review Models With Curated Training to Reduce Bias
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
User-generated reviews and question answers become verbose, repetitive, biased, and of varying quality, making it cumbersome to find meaningful information, and negativity bias often creeps into generative models trained on such content.
Innovation Solution
Fine-tune generative language models using reinforcement learning with a general corpus of user reviews and a vetted dataset to generate consensus reviews and answers, mitigating negativity bias and toxicity, and using bandit approaches to refine models based on user feedback.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If generative models are trained on user-generated reviews, then the model can learn from real user feedback, but the model inherits negativity bias and toxicity from the training data
Solution Approach 1:
The patent introduces an intermediary curating system that acts as a mediator between user-generated reviews and the training data. This system filters and selects only high-quality, helpful reviews based on multiple criteria (helpfulness votes, review length, relevance), thereby preventing the direct transmission of negativity bias and toxicity from raw user reviews to the model training data.
Solution Approach 2:
The patent extracts only the beneficial elements from user-generated reviews by implementing a curating process that identifies and isolates high-quality reviews. The extraction process removes harmful elements (negativity bias, toxicity, verbosity) while retaining useful information, creating a purified training dataset that maintains adaptability without inheriting harmful characteristics.
2Loss of information
If user reviews are used directly for training, then the model captures authentic user perspectives, but the reviews become verbose and repetitive making it cumbersome to find meaningful information
Solution Approach 1:
The curating system extracts only the essential and meaningful portions of user reviews by filtering out verbose and repetitive content. It identifies key information through multiple criteria including review length, helpfulness votes, and relevance, thereby maintaining authentic user perspectives while eliminating unnecessary verbosity that hinders information retrieval.
Solution Approach 2:
The patent applies different quality standards to different portions of review data. Rather than treating all reviews uniformly, the system evaluates each review locally based on specific criteria (helpfulness, length, relevance) and selectively includes only those portions that meet high-quality thresholds, thereby optimizing information density while preserving authentic perspectives.
3Adaptability or versatility
If all user reviews are included in training data, then the model covers diverse user opinions, but the quality of training data varies significantly reducing model performance
Solution Approach 1:
The patent changes the parameters used to evaluate and select training data by implementing multiple quality criteria (helpfulness votes, review length thresholds, relevance checks). This multi-parameter filtering approach transforms the raw, variable-quality review data into a standardized high-quality dataset that maintains diversity of opinions while ensuring consistent quality standards across all training examples.
Data Source
AI summary
Devices and techniques are generally described for generating content reviews using language models. In some examples, first text data may be determined for a first item. The first text data may include a plurality of user reviews of the first item. In various examples, the first text data may be input into a generative language model. In some cases, the generative language model may generate second text data including a first consensus review for the first item. In some examples, the second text data may be displayed on a graphical user interface in association with a listing for the first item.