Neuro-Symbolic Federated Research Artifact Search

Scientific research artifacts such as datasets, software or ontologies are essential components of scientific discovery. Yet, the growing volume of such artifacts requires more efficient and relevant search and retrieval systems. We present a neuro-symbolic approach for federated research artifact search, specifically for datasets and software metadata over Resodate and Wikidata. Integrated into the ORKG ASK platform, our system processes user queries through linguistic analysis to extract key terms. These key terms are then used to retrieve and recommend relevant research artifacts from federated sources, ensuring precise and contextually relevant metadata discovery. To further enhance retrieval accuracy, we employ a ranking mechanism that organizes research artifacts based on each user query’s structure and morphological features. We evaluate various key-term extraction methods and ranking approaches, integrating both symbolic and neural techniques. We rigorously evaluate the key-term extraction using Precision, Recall, and F1-score, and assess the re-ranking effectiveness by comparing with human rankings through correlation metrics and LLM-based evaluations. Our experiments show that symbolic methods outperform the neural approach regarding accuracy and response time. As a result, our system offers users more effective and efficient research artifact recommendations.