An Expert-independent Artificial Intelligence Method for Extracting and Prioritizing Decision Factors From Large-scale Textual Corpora
This paper introduces FEDRA, an autonomous framework that leverages transformer-based embeddings, clustering, and graph analysis to extract and prioritize decision factors from large-scale scientific text without expert intervention, demonstrating high accuracy and semantic consistency in a wastewater treatment site selection case study.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Every day, scientists, engineers, and policymakers face a mountain of written information. Thousands of research papers, technical reports, and policy documents are published constantly, each containing pieces of the puzzle needed to make difficult choices. For decades, the standard way to turn this mountain of text into a clear list of priorities has been to ask human experts. Groups of specialists would meet, read the literature, and vote on what matters most, often using structured surveys or lengthy discussions to reach a consensus. While this approach has served us well, it is slow, expensive, and relies entirely on the specific people chosen to participate. If the experts change, the results often change too. Furthermore, as the volume of available information grows faster than any team of humans can read, this traditional method is beginning to hit a wall.
A new approach is emerging that attempts to bypass the human bottleneck entirely. Instead of asking people to decide what is important, this method asks a computer to read the documents and find the patterns for itself. The core idea relies on two modern capabilities. First, computers can now understand the meaning of sentences, not just the words they contain, by converting text into mathematical representations of context. Second, they can map how different ideas appear together, revealing which concepts are naturally linked in the minds of researchers. By combining these tools, it is now possible to build a system that extracts the most critical factors for a decision directly from the text, without a single human expert needing to cast a vote or fill out a form.
Iman Mosaddegh, an independent researcher in artificial intelligence, has developed a framework called FEDRA to test this idea. The goal was to create a system that could automatically read a large collection of scientific literature, identify the key factors that influence a specific decision, and rank them by importance, all without human intervention. To see if this was possible, the researcher applied the system to a complex real-world problem: choosing the best location for a wastewater treatment plant. This is a decision that involves balancing many competing concerns, from the distance to rivers and roads to the cost of land and the needs of local communities. The researcher gathered twenty-five scientific documents published between 2012 and 2026, covering a total of over 121,000 words. This collection served as the raw material for the computer to analyze.
The process began by feeding the full text of these documents into the system. The computer broke the text down into sentences and paragraphs, then used advanced language models to understand the meaning behind the words. Instead of looking for specific keywords, the system looked for clusters of ideas that appeared together frequently. It identified twenty-eight distinct factors that the literature discussed as important for site selection. These factors were not pre-programmed by the researcher; they emerged naturally from the data. The system then organized these factors into four logical groups: accessibility and demand, environmental and hydrological conditions, engineering and infrastructure constraints, and socioeconomic considerations. This grouping happened automatically, based on how the factors were discussed in relation to one another within the text.
Once the factors were identified, the system had to decide which ones were the most important. It did this by building a network map of the factors. In this map, every factor was a point, and lines connected factors that appeared together in the same sentences or paragraphs. The system then analyzed this network to see which factors were the most central. Factors that were connected to many other important factors received a higher rank, much like a popular person in a social network who is connected to many other influential people. This method allowed the system to prioritize the factors based on their structural importance within the literature, rather than just counting how often a word appeared. The top-ranked factor identified by the system was proximity to water bodies, followed closely by distance from roads and population density. These results aligned closely with what human experts have traditionally concluded, suggesting the computer had successfully learned the same lessons from the text.
To verify the results, the researcher compared the computer's list against a benchmark created by a large language model, which acted as a second, independent reader of the same documents. The comparison showed a strong agreement between the two methods. The computer successfully identified twenty-eight out of the twenty-nine key factors that the benchmark model found, a success rate of over ninety-six percent. The ranking of these factors also matched closely, with a high statistical correlation indicating that the computer and the benchmark model agreed on which factors were most critical. The system also demonstrated that the factors it found were semantically consistent, meaning the groups of factors it created made logical sense together. While the system did miss one factor related to cultural heritage and aesthetics, its overall performance was robust enough to show that it could reconstruct a meaningful decision space from unstructured text.
The study suggests that it is possible to automate the discovery of decision factors, offering a scalable alternative to traditional expert panels. The framework proved capable of handling a diverse set of documents and extracting a structured, interpretable list of priorities without any human scoring or predefined categories. However, the researcher notes that the system is not entirely free of human influence. The quality of the output depends on the quality of the documents fed into it, and the researcher still had to choose which documents to include and how to set the system's parameters. The system does not replace the need for human judgment in the final decision-making process, but it offers a powerful tool for gathering and organizing the evidence that informs those decisions. By turning a massive library of text into a clear, ranked list of factors, this approach could help organizations make faster, more consistent decisions in fields ranging from engineering to public policy, provided they have access to a solid body of written knowledge to analyze.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.