Meta-predictive ensembles of LLM pipelines for automated named entity recognition in aerospace software requirements

In safety-critical aerospace systems, errors in requirements engineering can lead to significant operational risks, which makes accurate information extraction essential. Named Entity Recognition (NER) plays a crucial role in this process by supporting tasks such as requirements classification and automated test case generation. Despite the emergence of diverse language model–based predictive pipelines across multiple domains for NER, the aerospace software development field has so far witnessed only limited exploration of the capabilities of foundational large language models (LLMs). This study investigates the effectiveness of foundational LLMs, recurrent architecture–enhanced LLMs, and generative models under domain-specific constraints. However, single-model approaches struggle with stability, domain shifts, and difficulty in capturing long-context dependencies, which motivates the need for more robust solutions. To address this, the study proposes a novel meta-predictive pipeline that integrates heterogeneous models for improved robustness. To develop a robust and precise meta predictor, it benchmarks the performance of 15 foundational LLMs, 30 recurrently enhanced LLMs, and 4 generative language models, including DeepSeek, ChatGPT, LLaMA, and Gemini. Finally, 12 distinct meta-predictors were developed based on foundational LLMs and recurrent architectures. The best performing meta predictor on the public AeroBERT-NER benchmark dataset achieved 94.73%, 93.59%, and 94.17% precision, recall, and F1-score, respectively. The key findings demonstrate the potential of the proposed approach to improve reliability and scalability in aerospace requirements engineering workflows.