TACTICIAN - AI‑Driven Knowledge Extraction for ESA Space Science Missions

Abstract

Introduction: ESA’s missions produce vast scientific data and thousands of publications annually. Identifying those that truly analyze mission data is essential but manual review is unfeasible. TACTICIAN uses NLP, ML, DL, and LLMs to automate extraction and classification of mission-related knowledge. Methodology and Relation Extraction: The TACTICIAN system presents an integrated approach for converting unstructured scientific literature into structured knowledge that is directly connected to specific space science missions. By leveraging a combination of syntactic and semantic Natural Language Processing (NLP) techniques, the framework effectively extracts and organizes mission-relevant information. This process includes several key components: Part-ofSpeech (POS) tagging, which identifies grammatical structures; dependency parsing to elucidate relationships between words; named-entity recognition to pinpoint mission-specific instruments, targets, and observation identifiers; temporal expression extraction to capture relevant timeframes; and entity-based summarization that synthesizes core findings. Collectively, these methods enable TACTICIAN to detect crucial mission characteristics such as instruments, targets, wavelengths, observation identifiers, and predicates indicating how scientific data is utilized, thus facilitating the transformation of complex scientific texts into actionable, mission-linked datasets. Scalable LLM Classification and Reflection: Recent advancements in the TACTICIAN project emphasize the use of classification pipelines built upon Large Language Models (LLMs). These pipelines are specifically designed to operate without the need for missionspecific training, enabling robust performance even when applied to new missions or novel datasets. By employing zero-shot and few-shot prompting strategies, the system can effectively classify publications and extract relevant information without extensive prior training. The integration of automatic summarization and structured JSON outputs ensures that the extracted knowledge is organized and easily accessible, while delivering results that are on par with the most effective Deep Learning (DL) models. A notable feature of this approach is LLM reflection. In this process, the model initially generates a prediction based on the input, then formulates a probing question to test the validity or depth of its response and finally reassesses its answer considering this selfquestioning. This reflective methodology enhances both the accuracy and the explainability of the classification outcomes, making the pipeline not only powerful but also transparent and trustworthy for missionlinked literature analysis. Conclusion: The TACTICIAN system leverages advanced Natural Language Processing, Machine Learning, Deep Learning, and Large Language Model technologies to reliably classify space science literature, accurately extract mission-specific knowledge,and establish connections between publications and observational archives. Traditional ML and DL approaches achieve high accuracy in organizing missionlinked information, while LLM-based methods offer scalable, mission-agnostic solutions ideal for analyzing future space mission literature. These capabilities directly support the European Space Agency’s goals by enabling scientific impact assessment, promoting open-science reproducibility, improving intelligent archive services, and laying the groundwork for expanding multimodal and agentic AI applications in the space domain. Acknowledgments: This work has received funding from the European Space Agency (ESA Contract No 4000128429/19/ES/JD).