Automated Literature Review via Trained Language Model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systematic literature reviews are laborious, error-prone, and time-consuming due to the manual nature of the screening process, which can lead to delays in incorporating new studies and requires significant human resources, while existing AI and ML methods are inaccurate and require extensive maintenance and training.
Innovation Solution
A computer-implemented method and system for automated systematic literature review using a trained language model that processes search results based on inclusion and exclusion criteria, generating selection scores to efficiently prescreen publications, reducing the need for multiple models and improving accuracy and resource efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual literature review process is used, then review accuracy can be maintained through human judgment, but the process becomes laborious and time-consuming
Solution Approach 1:
A trained language model is introduced as an intermediary between the large volume of search results and human reviewers. The model automatically prescreens publications by formulating questions based on inclusion/exclusion criteria, generating selection scores, and ranking results. This intermediary handles the time-consuming initial filtering task, allowing human reviewers to focus on evaluating the smaller subset of prescreened results, thereby resolving the contradiction between maintaining review accuracy and reducing prescreening time.
2Productivity
If existing AI and ML methods are used for automated review, then processing speed increases, but accuracy decreases and maintenance requirements increase
Solution Approach 1:
The patent fine-tunes a pre-trained language model on a question-and-answer task with specific parameters optimized for literature review. By adjusting the model's parameters through fine-tuning on domain-specific data and criteria, the system achieves both high processing speed and high accuracy. This resolves the contradiction by demonstrating that proper parameter optimization can simultaneously improve productivity and reliability, unlike generic AI methods.
Solution Approach 2:
The system incorporates feedback mechanisms where the language model generates selection scores that are evaluated and refined. The model learns from the evaluation process, allowing continuous improvement of accuracy while maintaining automated processing speed. This feedback loop enables the system to achieve high accuracy without requiring extensive manual maintenance.
3Adaptability or versatility
If multiple AI models are used for different review aspects, then comprehensive coverage is achieved, but system complexity and resource requirements increase
Solution Approach 1:
The patent employs a single multi-functional language model that can handle multiple aspects of literature review through one unified system. The model formulates questions, evaluates publications against multiple inclusion/exclusion criteria, generates selection scores, and ranks results all through one model rather than requiring separate models for each task. This universal approach achieves comprehensive review coverage while significantly reducing system complexity and resource requirements compared to using multiple specialized models.
Data Source
AI summary
Publication pre-screening may include the use of a trained model. A trained language model may be fine-tuned on a question-and-answer task and may be configured to receive a question that includes inclusion and exclusion criteria for a publication. The question may be formulated to include context information such as a title and abstract of the publication. An output of the model may be used to determine a selection of the publication.