Formalization of requirements for software and information systems using natural language processing methods

Abstract

The extraction and formalization of requirements from technical specification texts remains a poorly automated stage of the software and information system life cycle: a substantial share of requirements is expressed in natural language, while their translation into a verifiable formal form is performed manually and depends on the analyst's qualification. This paper presents an author's model of multi-loop requirement extraction and formalization that combines domain production rules, statistical segmentation and a fine-tuned language model, followed by verification against the requirement quality criteria of ISO/IEC/IEEE 29148. The model was tested on a model corpus of 214 technical specification texts containing 4,187 requirements of five types. The weighted F1 score of extraction reached 0.856, with a formalization completeness of 0.86 and a share of unambiguous formalizations of 0.88. A comparison of four approaches showed that a fine-tuned transformer yields the highest extraction accuracy (F1 0.869) but is inferior to the hybrid model in the unambiguity of formalization (0.77 against 0.88), so that the advantage in accuracy does not translate into an advantage in the verifiability of the result. The largest share of errors falls on the omission of implicit requirements and incorrect segmentation of compound formulations. A stable saturation of quality occurs at a training corpus volume of about 80%. The obtained relations show that the bottleneck of formalization shifts from recognition accuracy toward ensuring unambiguity and verifiability, which entails the priority of post-processing over increasing the power of the recognizing component.