An Agentic Framework Using Rules and LLMs for Embedding and Annotating Descriptive Document Layouts: A Plant Science Use Case
This paper presents a modular, agent-based framework that combines rule-based parsing and large language models to robustly extract and enrich botanical trait annotations from descriptive document layouts, successfully processing over 55,000 traits across nearly 5,000 plant species.
Original paper licensed under CC BY 4.0 (http://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
For centuries, botanists have relied on thick, printed volumes to describe the world's plants. These books, often filled with dense paragraphs of text, hold the keys to understanding how a species lives, where it grows, and what it looks like. Yet, for computers to use this knowledge, the information must be converted from human-readable pages into structured data that machines can search and analyze. The challenge lies in the sheer volume and complexity of these texts, which are written in many languages and contain subtle descriptions of plant features, known as traits. Extracting this data manually is slow and prone to error, while traditional computer methods often struggle to understand the nuance of scientific language or the layout of old, scanned pages. To solve this, researchers have begun building systems that combine strict logical rules with the flexible reasoning of artificial intelligence, aiming to turn libraries of botanical descriptions into vast, searchable databases.
In a recent study, a team of researchers developed a new system designed to read these botanical texts and automatically pull out specific details about plant characteristics. The process begins with digital scans of old books, which are often just images of text. The system first uses optical character recognition, a technology that turns those images into actual letters and words a computer can read. Once the text is readable, the system organizes it by grouping descriptions under the correct plant names. This step is crucial because a single book might describe thousands of different species, and the computer must know exactly where one plant's description ends and another begins. The researchers found that simply reading the text was not enough; the system needed to understand the structure of the page to separate the different sections accurately.
To extract the actual information, the team built a pipeline that acts like a team of specialized workers. The first group of workers follows strict, pre-written rules to find specific patterns. For example, if the text mentions "leaves" followed by a measurement like "4 centimeters," the system records that the plant has leaves of that size. These rules are precise and reliable, but they can only find what they are explicitly told to look for. To catch more details, the system employs a second group of workers powered by large language models, which are advanced artificial intelligence programs trained on vast amounts of human writing. These AI agents act as virtual experts who can suggest new words and phrases that describe plant parts, helping the system recognize descriptions it might have missed with rules alone. This combination allows the system to be both accurate and adaptable, filling in gaps without losing the clarity of the original data.
The researchers tested this system on three large collections of botanical texts from New Caledonia, Senegal, and Cameroon. The results showed that the system successfully extracted nearly 56,000 specific trait annotations across almost 5,000 different plant species. On average, the system identified about nine distinct traits for each plant it analyzed. When they added the AI-powered vocabulary expansion to the strict rule-based system, the total number of annotations increased by nearly 60 percent, covering three-quarters of the traits they were looking for. This improvement demonstrated that while the rules provided a solid foundation, the AI agents significantly expanded the system's ability to understand the diverse ways botanists describe plants.
The study also revealed that the system is robust, meaning it performs consistently even when the quality of the scanned text varies. The researchers tested different methods for converting the images to text and found that while some methods were slightly better at recognizing plant names, the overall number of traits extracted remained stable regardless of the method used. This suggests that the pipeline is reliable enough to handle the messy reality of historical documents. However, the system is not fully automatic; it still requires human experts to review the suggestions made by the AI agents before they are added to the final rules. This human oversight ensures that the data remains trustworthy and that the system does not invent facts, a common risk when using artificial intelligence alone.
Ultimately, this work provides a powerful tool for organizing the world's botanical knowledge. By turning unstructured text into organized data, the system helps scientists build better databases for studying biodiversity and plant evolution. The approach proves that combining the precision of human-designed rules with the flexible reasoning of artificial intelligence can unlock information hidden in millions of pages of scientific literature. The researchers made their code and data available for others to use, ensuring that this method can be applied to other fields where complex text needs to be understood and organized.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.