AI-based Text Generation for Semantic Search Robustness : Application to Defence

Abstract

One of the biggest challenges in successfully applying Artificial Intelligence (AI) in the Defense sector is the availability of trustful domain specific data to train AI models on. These data have to be generated and collected from the real world or acquired through realistic scenarios simulations and validated by operation specialists or domain experts. In real world applications, most of the time these data are classified and difficult to access. Then only a handful coming either from unclassified documents or simulation / realistic scenarios can be made available. In this article, we discuss how Generative AI can be used to generate intelligence-oriented textual data that are semantically similar to a “ground truth” database. The methodology is applied in the frame of the EDIDP AI4DEF project, focusing on one of the use cases, Request for Information (RFI) semantic similarity detection in a database. We expose how a limited corpus has been enriched with noisy AI-generated data. The performances and the robustness of the AI model have been monitored to be kept similar before and after the data augmentation, while a human-in-the loop qualifies the AI-generated data.