AUTOMATED DETECTION AND CLASSIFICATION OF AMBIGUITIES IN SOFTWARE REQUIREMENTS OF IT PROJECTS USING LARGE LANGUAGE MODELS AND RAG TECHNOLOGIES

Abstract

The problem of ensuring the quality of requirements in the early stages of the development life cycle is of critical importance for the success of software projects, since the presence of linguistic and semantic ambiguities in specifications is one of the main causes of defects, budget overruns and disruptions in implementation deadlines. Manual review of documentation is an excessively laborious and subjective process, and traditional automated rule-based tools often generate a high level of false positives due to the inability to understand the context. The use of large language models opens up new opportunities for semantic text analysis, but their direct application is complicated by the tendency to hallucinations and ignoring specific project terminology.

The article proposes a hybrid method for detecting ambiguities in software requirements, which is based on the integration of large language models with the technology of search-addition generation (RAG). The proposed approach is based on a two-agent architecture, which involves the sequential interaction of an identifier agent for initial filtering of requirements and a classifier agent for their deep semantic typing according to the IEEE 830 standard. The method includes such stages as text segmentation, dynamic enrichment of queries with context from the project knowledge base, application of the Chain-of-Thought strategy for generating explanations, and structural parsing of results.

The experiments were conducted on a specialized dataset from the telecommunications domain containing 1983 requirements. The results obtained demonstrated the superiority of the developed hybrid method based on the Flan-T5-Large model over the basic approaches (Zero-Shot and Few-Shot): the method provides Precision 87%, Recall completeness 91%, and F1-score 89%. Additionally, the effectiveness of using RAG to reduce the number of false positives on highly specialized technical vocabulary is confirmed and the system's ability to provide interpreted explanations of detected defects is demonstrated. The results prove that the integration of context-oriented LLM agents significantly increases the level of automation and reliability of the requirements audit process in modern development environments.