A Hybrid Approach to Extraction of Knowledge From Scientific Texts Based on Large Language Models and Domain Dictionaries

The vast information landscape of the Internet constitute a significant challenge for extracting valuable content. The lack of standardized data models and structures necessitates ad hoc solutions, often requiring expert knowledge that developers may lack. While Large Language Models (LLMs) hold promise for addressing this challenge, their susceptibility to AI hallucinations and inaccuracies necessitates ongoing research. This paper introduces a hybrid approach for extracting information on computational methods that combines domain-specific dictionaries with LLMs in order to improve the accuracy of method categorization. Our system incorporates explainability, allowing users to understand the reasoning behind method assignments. Furthermore, user-driven training is facilitated by allowing users to select theories and highlight relevant keywords, enhancing learning capabilities of the system. Its implementation demonstrates the effectiveness of this approach, achieving an impressive Fl-score of up to 90.3 %. This research contributes to the ongoing effort to develop robust and accurate knowledge extraction systems for navigating the ever-expanding landscape of online information.