Optimizing Named Entity Recognition for Improving Logical Formulae Abstraction from Technical Requirements Documents

Requirements engineering involves obtaining a complete and consistent set of requirements for a particular system. Requirement engineers often formulate requirements in a textual form where automated proving consistency and completeness is impossible. Converting textual requirements into logical formulae would enable analysis and reasoning tasks. In this work, we contribute to converting textual requirements into logical sentences. In particular, we focus on the named entity recognition task on industrial requirements documents to improve semantic abstraction using logical formalisms. We found that using general-purpose models for named entity recognition is not working well for specialized domains. Hence, we focused on retraining such models and investigated the ratio of domain-specific versus general-purpose data for retraining. Our results demonstrate significant improvements in F1-Scores compared to the respective pre-trained baseline models when performing domain-specific tasks.