Abstract:Requirements Engineering (RE) is a critical phase in the software development process that generates requirements specifications from stakeholders' needs. Recently, deep learning techniques have been successful in several RE tasks. However, obtaining high-quality requirements specifications requires collaboration across multiple tasks and roles. In this paper, we propose an innovative framework called MARE, which leverages collaboration among large language models (LLMs) throughout the entire RE process. MARE divides the RE process into four tasks: elicitation, modeling, verification, and specification. Each task is conducted by engaging one or two specific agents and each agent can conduct several actions. MARE has five agents and nine actions. To facilitate collaboration between agents, MARE has designed a workspace for agents to upload their generated intermediate requirements artifacts and obtain the information they need. We conduct experiments on five public cases, one dataset, and four new cases created by this work. We compared MARE with three baselines using three widely used metrics for the generated requirements models. Experimental results show that MARE can generate more correct requirements models and outperform the state-of-the-art approaches by 15.4%. For the generated requirements specifications, we conduct a human evaluation in three aspects and provide insights about the quality
What problem does this paper attempt to address?
### What problems does this paper attempt to solve?
This paper aims to solve the problems of automation and collaboration in the Requirements Engineering (RE) process. Specifically, the paper proposes a multi - agent collaboration framework MARE (Multi - Agent collaboration for Requirements Engineering) to achieve collaboration between different tasks and roles in the RE process, thereby generating high - quality requirements specifications.
#### Specific problem description:
1. **Complexity and multi - task nature of requirements engineering**:
- Requirements engineering is a crucial stage in the software development process, involving generating requirements specifications from the needs of stakeholders.
- In order to obtain high - quality requirements specifications, collaboration among multiple tasks and roles is required, including tasks such as requirements elicitation, modeling, verification, and specification.
2. **Limitations of existing methods**:
- Although deep - learning techniques have achieved success in some requirements engineering tasks, automating only a few tasks limits the improvement of overall effectiveness and efficiency.
- Most of the existing research focuses on using large - language models (LLMs) to handle specific requirements tasks, without fully exploiting the collaborative potential of LLMs between different tasks.
3. **Need for multi - agent collaboration**:
- Research shows that multi - agent collaboration mechanisms can improve the effectiveness of handling complex tasks. Therefore, designing a framework that can promote collaboration between different agents is crucial for improving the quality of requirements engineering.
#### Main contributions of the MARE framework:
- **End - to - end automation**: MARE realizes the automation of the entire requirements engineering process, from the initial requirements concept to the final generation of requirements specifications.
- **Multi - agent collaboration**: MARE completes the four main tasks (elicitation, modeling, verification, and specification) of requirements engineering through five agents (stakeholders, collectors, modelers, inspectors, and document writers) and nine actions.
- **Shared workspace**: MARE provides a shared workspace for storing and sharing intermediate requirements artifacts, promoting collaboration between agents.
### Experimental results and evaluation:
1. **Requirements modeling performance**:
- Experiments were carried out on five public cases, one public dataset, and four newly created cases, using Precision, Recall, and F1 - score as evaluation metrics.
- The experimental results show that MARE significantly outperforms the existing state - of - the - art methods in requirements modeling, with an average F1 - score improvement of 15.4%.
2. **Requirements specification quality**:
- A manual evaluation was carried out on the generated requirements specifications, scoring from three aspects: correctness, completeness, and consistency.
- The results show that the requirements specifications generated by MARE perform excellently in these aspects, with average scores of 0.98, 1.92, and 1.98 respectively.
3. **Ablation study**:
- The performance of a single LLM and MARE in the requirements modeling task was compared, proving the effectiveness of the multi - agent collaboration framework.
In conclusion, this paper solves the problems of multi - task collaboration and automation in the requirements engineering process by proposing the MARE framework, significantly improving the quality and generation efficiency of requirements specifications.