Quinex AI extracts numerical data from articles with 98% accuracy

Quinex AI Tool Extracts Numerical Data from Scientific Articles with 98% Accuracy

Researchers from Forschungszentrum Jülich (Jülich, Germany) have introduced the Quinex (Quantitative Information Extraction) framework, designed for automatic extraction of quantitative data from scientific publications. The system, based on open language models, can find numerical values, units of measurement, and contextual information, transforming unstructured text into structured datasets. The accuracy of extracting numbers and units reaches 98% F1-score, making the tool promising for accelerating literature reviews in energy, climatology, materials science, and other fields.

How Quinex Works

The framework processes the text of scientific articles and extracts quantitative parameters: temperatures, efficiency, emissions, cost, medical indicators, etc. The system not only finds numbers but also determines which object or phenomenon they refer to, links units of measurement, records temporal and spatial coordinates, and the method of data acquisition. For example, the phrase "in 2025, efficiency will range from 63% to 71%" is transformed into a record with the parameter "efficiency," a range of values, a year, and a source reference. This allows researchers to quickly collect and compare data from hundreds and thousands of articles.

Open Models and Transparency

A key feature of Quinex is the use of open and compact language models instead of closed commercial solutions. This approach simplifies verification, refinement, and deployment of the system without expensive infrastructure. The developers emphasize that the tool does not "hallucinate" numbers: values are extracted directly from the text, and errors are only possible when interpreting context, when relationships between data are scattered across different parts of the article.

Test Results

Quinex has been tested on thousands of abstracts from various scientific fields. In addition to 98% accuracy for numbers and units, the system achieved 87% F1 for classifying quantitatively described properties and 82% for entities. Test sets included data on electricity production costs, maximum human oxygen consumption, earthquake magnitudes, and bandgap width in photovoltaic materials.

What Matters

• Quinex automates the extraction of numerical data from scientific articles, reducing time for literature reviews.
• The system uses open language models, ensuring transparency and accessibility.
• The accuracy of extracting numbers and units of measurement is 98% F1.
• The tool is intended for auxiliary use: final conclusions remain with the researcher.
• The project is released as open source for adaptation to specific disciplines.

Context and Prospects

The flow of scientific publications is growing faster than the capacity for manual analysis. Quinex addresses the "bottleneck" problem in research where comparing quantitative parameters from multiple sources is required. The developers plan to expand the tool with industry-specific datasets and new models to improve accuracy in specific areas such as energy, chemistry, and biomedicine.

FAQ

Question: What types of data can Quinex extract?

Answer: The system processes any numerical values linked to units of measurement: temperatures, efficiency, emissions, cost, medical indicators, and other quantitative parameters.

Question: Can Quinex be used without special computing resources?

Answer: Yes, the tool is built on open and compact models, allowing it to run without expensive infrastructure.

Question: Does Quinex replace the work of a researcher?

Answer: No, the system is an auxiliary tool for data collection, while analysis and conclusions remain with the scientist.

— Editorial Team