The rapid growth of the low-altitude economy and the complexity of airborne software have drawn many novice engineers into a field that demands deep domain expertise and strict development processes. At Avicas Generic Technology Co., Ltd., we maintain a repository of internal engineering guidelines and standards, but its volume and complexity limit accessibility, especially for newcomers. We propose a domain-specific synthetic question-answering (QA) dataset construction framework to bridge this gap. The framework comprises Generation, Curation, and Evaluation phases, each incorporating semantic-oriented quality-control strategies. Using it, we built and assessed a dataset with 1241 QA pairs for airborne software development. Expert review found that over 80% of QA pairs were usable, and an auxiliary model was rated 97% as high quality. Fine-tuning a large language model (LLM) on this dataset improved downstream task accuracy to an average of 94%, an 8.5% gain over the baseline. We release $\mathbf{1 0 \%}$ sample for research use and share practical insights to support domain knowledge transfer and the creation of specialized LLMs.

Threat analysis is a pillar of security-by-design which plays an important role in the elicitation and refinement of security threats. In preparation for the analysis, a model of the system under analysis e.g., the Data Flow Diagram (DFD for short) is often created. Empirical measures of success are important for practitioners that are struggling to meet the current demands for expertise. But no previous work has investigated the role of these diagrams during the validation of identified security threats. This paper presents an experiment conducted with 98 students in two countries. We measured the impact of the DFD on the perceived and actual effectiveness of validating a list of identified security threats including both fabricated and actual threats. In presence of sequence diagrams, the participants perceived DFDs as more useful. However, when exposed to both a DFD and a sequence diagram, DFDs had no significant impact on the participants’ ability to validate security threats.

The research of machine learning for requirements engineering (ML4RE) has attracted more and more attention from researchers and practitioners. Although pioneering research has shown the potential of using ML techniques to improve RE practices, there lacks a systematic and comprehensive literature review in academia that integrates an industrial perspective. Specifically, none of the reviews available in ML4RE have considered the grey literature, which is primarily from practitioner origin and is more reflective of the real issues and challenges faced in practice. In this paper, we conduct a systematic survey of academic publications in ML4RE and complement it with the practitioners’ voices from Stack Overflow to complete a comprehensive literature review. Our research objective is to provide a comprehensive view of the current research progress in ML4RE, present the main questions and challenges faced in RE practice, understand the gap between research and practice, and provide our insights into how the RE academic domain can pragmatically develop in the future. We systematically investigated 207 academic papers on ML4RE from 2010 to 2022, along with 375 questions related to RE practices on Stack Overflow and their corresponding answers. Our analysis encompassed their trends, focused RE activities and tasks, employed solutions, and associated data. Finally, we conducted a joint analysis, contrasting the outcomes of both parts. Based on the statistical results from collected literature, we summarize an academic roadmap and analyse the disparities, offering research recommendations. Our suggestions include the development of intelligent question-answering assistants employing large language models, the integration of machine learning into industrial tools, and the promotion of collaboration between academia and industry. This study contributes by providing a holistic view of ML4RE, delineating disparities between research and practice, and proposing pragmatic suggestions to bridge the academia-industry gap.

Goal modeling plays an imperative role in early requirements engineering, which has been investigated for decades. There have been many studies that show the usefulness of requirements goal models. However, the establishment of goal models is typically done manually, which is time-consuming and has a steep learning curve. In this paper, we propose a semi-automatic framework for constructing iStar models, which is a well-known goal modeling language. Specifically, we first investigate the practical needs of iStar modelers on the automation of iStar modeling by holding interviews, based on which we propose an interactive and iterative modeling process. Our proposal takes advantage of human decisions and artificial intelligence algorithms, respectively, aiming at achieving low modeling costs while maintaining the quality of models. We then propose a hybrid approach for automatically extracting goal model snippets from requirements text, which implements the automatic tasks of our proposed process. The proposed method combines logical reasoning with deep learning techniques so as to unleash the power of domain knowledge to assist with automation tasks. We have performed a series of experiments for evaluation. The experimental results show that our method achieves the F1-measure of 90.34% for actor entity extraction, 93.14% for intention entity extraction, and 83.18% for actor relation extraction, which can efficiently establish high-quality goal models. The artifacts are available at Zenodo1.

iStar modeling is an effective measure for requirements analysis, and researchers have proposed nearly thirty different modeling tools for this purpose. There are two types of existing modeling tools, i.e., graphical and textual. However, either type has its limitations. Graphical modeling tools suffer from the scalability issue. Textual tools rely on visual models when modeling. To overcome the limitations, we have developed the BiStar, a template-based iStar modeling tool combining graphical and textual modeling. The two main features of our BiStar tool are as follows. First, BiStar adopts textual templates to support the batch addition of four common iStar model elements, including Actors, Intentions, Dependencies, and Refinements. Second, BiStar supports the templated creation of iStar models using text models. In addition, BiStar has full support for graphical iStar modeling as a basic feature of a modeling tool. BiStar automatically visualizes the model elements and templates added in the text as described above. After visualization, modelers can freely choose between graphical and textual modeling to continue modeling. Our BiStar tool thus combines graphical modeling with textual modeling.

[Context] Establishing requirements models is an effective way to analyze them, which is typically dealt with in a graphical manner (i.e., the drag-and-draw fashion). However, as the size of models increases, the scalability issue has become an unignorable challenge, hindering the practical adoption of requirements modeling approach. Some researchers have recently proposed and promoted textual modeling approaches, mitigating these issues of requirements modeling. [Objective] In this paper, we aim at evaluating the two modeling methods, i.e., a graphical modeling method VS. a textual modeling method. In particular, we apply these two methods to iStar modeling language, which has been widely recognized as an effective means to model and analyze requirements. [Methods] We have systematically designed and conducted a controlled experiment with 38 participants to compare two iStar modeling methods (graphical and textual) using two corresponding modeling tools (piStar and T-Star). The experimental results reveal that the numbers of iStar model nodes and relationships built by the participants had no significant difference, regardless of the modeling method adopted. [Conclusions] First, the results show that the textual modeling method is as usable as the graphical modeling method when creating iStar models. Second, we have identified a number of issues that contribute to improving the utility and practicality of the iStar modeling method.

iStar framework is an effective means for modeling and analyzing goals and interactions among social agents. Most of existing iStar modeling tools build iStar models via a graphical manner, which has a steep learning curve and suffer from scalability issues. Especially when dealing with large-scale models in industrial settings, it is very time consuming to play with the layout of models. In this paper, we present a tool which allows users to build iStar models from textual descriptions of systems. And the established models are then visualized with reasonable layouts. In particular, our tool supports both SD (Strategic Dependency) view and SR (Strategic Rationale) view of iStar models.