Previous work has investigated the adequacy of LLMs to detect inconsistencies in requirements documents, but has also shown their limitations with real case studies. In this paper, we propose a hybrid approach, which exploits traditional clustering techniques to help LLMs focus on potential inconsistencies. The approach was evaluated using a large security requirements document from the RE Open Data Initiative, with injected inconsistencies. Results show that combining LLM-based detection with rule-based clustering enhances both precision and recall.

With the rapid advancement of tools based on Artificial Intelligence, it is interesting to assess their usefulness in requirements engineering. In early experiments, we have seen that ChatGPT can detect inconsistency defects in natural language (NL) requirements, that traditional NLP tools cannot identify or can identify with difficulties even after domain-focused training. This study is devoted to specifically measuring the performance of ChatGPT in finding inconsistency in requirements. Positive results in this respect could lead to the use of ChatGPT to complement existing requirements analysis tools to automatically detect this important quality criterion. For this purpose, we consider GPT-3.5, the Generative Pretrained Transformer language model developed by OpenAI. We evaluate its ability to detect inconsistency by comparing its predictions with those obtained from expert judgments by students with a proven knowledge of RE issues on a few example requirements documents.

The engineering of software product lines begins with the identification of the possible variation points. To this aim, natural language (NL) requirement documents can be used as a source from which variability-relevant information can be elicited. In this paper, we propose to identify variability issues as a subset of the ambiguity defects found in NL requirement documents. To validate the proposal, we single out ambiguities using an available NL analysis tool, QuARS, and we classify the ambiguities returned by the tool by distinguishing among false positives, real ambiguities, and variation points, by independent analysis and successive agreement phase. We consider three different sets of requirements and collect the data that come from the analysis performed.

We introduce a temporal logic to reason on global applications. First, we define a modal logic for localities that embeds the local theories of each component into a theory of the distributed states of the system. We provide the logic with a sound and complete axiomatization. Then, we extend the logic with a temporal operator. The contribution is that it is possible to reason about properties that involve several components in a natural way, even in the absence of a global clock, as required in an asynchronous setting.