An analytical approach to named entity recognition for military aerospace intelligence

Named Entity Recognition (NER) models have been used in various contexts to automate the analysis of textual data, supporting tasks such as text categorization and relationship extraction across entities. While these models are typically trained on large corpora, their efficacy should be evaluated in specific domains where texts include specialized jargon and terminology. We assess the performance of NER models on Spanish-language documents from the Colombian Aerospace Force, focusing on the analysis of textual reports. Transformer and convolutional architectures were applied to three datasets. Our findings suggest that these models, when used off-the-shelf, face limitations, particularly in recognizing traditional words, military jargon, and compound entities. In Colombia, where names often include two given names and two surnames, and place names span multiple words, the models struggled to accurately parse and identify these linguistic patterns. However, after fine-tuning, their performance improved significantly, enabling effective extraction of valuable information from texts containing aerospace terms and military jargon in Spanish. As part of this study, we release six fine-tuned NER models and two of the datasets used in our experiments. Further research may yield additional optimizations, particularly for aerospace intelligence applications in Latin America. • Evaluate the accuracy of entity extraction in Spanish military aerospace documents. • Identify model limitations in recognizing compound and contextual entities. • Enhance decision-making by fine-tuning models for local language structure. • Demonstrate how tailored models improve military text analysis outcomes. • Recommend strategies for adapting language models to national defense needs.