NICE: Non-Functional Requirements Identification, Classification, and Explanation Using Small Language Models

Abstract

Accurate identification and classification of Non-Functional Requirements (NFRs) is essential for informed architectural decision-making and maintaining software quality. Numerous language model-based techniques have been proposed for NFR identification and classification. However, understanding the reasoning behind the classification outputs of these techniques remains challenging. Rationales for the classification outputs of language models enhance comprehension, aid in debugging the models, and build confidence in the classification outputs. In this paper, we present NICE, a tool for NFR Identification, Classification, and Explanation. Using an industrial requirements dataset, we generated explanations in natural language using the GPT-4o large language model (LLM). We then fine-tuned small language models (SLMs), including T5, Llama 3.1, and Phi 3, with these LLM-generated explanations to identify and classify NFRs and to explain their classification outputs. We evaluated NICE using standard evaluation metrics such as F1-score and human evaluation to assess the quality of the generated explanations. Among the models tested, T5 produced explanations of quality comparable to Llama 3.1 and Phi 3 but achieved the highest average F1-score of 0.90 in multi-label NFR classification on the industrial requirements dataset. Furthermore, to evaluate the effectiveness of NICE, a survey was conducted with 20 requirements analysts and software developers. NICE is currently deployed as a part of the Knowledge-assisted Requirements Evolution (K-RE) framework developed by a large IT vendor organization.