Towards a Software Reference Architecture for Natural Language Processing Tools in Requirements Engineering

1. Problem Statement

Natural Language Processing (NLP) tools are increasingly used to automate Requirements Engineering (RE) tasks (e.g., elicitation, classification, validation). However, the current landscape of NLP4RE tools suffers from critical systemic issues:

• Monolithic Development: Tools are built from scratch as isolated, tightly coupled systems, leading to the redundant re-implementation of common steps (e.g., input parsing, data formatting, visualization).
• Lack of Longevity: Most tools are academic prototypes that are abandoned shortly after publication, lacking long-term maintenance.
• Poor Interoperability & Reusability: The absence of architectural guidance prevents the reuse of existing modules, forcing developers to rebuild entire pipelines for new use cases.
• Incomparability: Diverse architectures make it difficult to benchmark tools against common standards or compare specific algorithmic improvements in isolation.
• Documentation Gaps: Documentation is often superficial, hindering adoption and industrial application.

2. Proposed Vision

The authors propose a paradigm shift from isolated monolithic tools to an ecosystem of interoperable, reusable modules. This vision aims to:

• Improve Reuse: Encapsulate common pipeline steps (e.g., CSV parsing) into modular components that can be shared across different tools.
• Enhance Applicability: Allow tools to adapt to different input formats (e.g., swapping a CSV parser for a DOCX parser) without rewriting the core logic, provided a common internal representation is used.
• Facilitate Comparability: Enable the isolation and benchmarking of specific modules (e.g., classification algorithms) while keeping the rest of the pipeline constant.
• Ensure Maintenance: Distribute maintenance responsibilities across the community, increasing the longevity of widely used components.

3. Methodology

The paper outlines a research roadmap to develop a Software Reference Architecture (SRA) for NLP4RE tools, following the ProSA-RA method (Nakagawa et al., 2014). The methodology consists of four steps:

1. Information Source Investigation: Gathering data from publications, existing software systems, reference models, ontologies, and stakeholders.
2. Architectural Analysis: Eliciting three types of requirements:
   • System Requirements (SR): Functional and non-functional needs.
   • Architectural Requirements (AR): Abstractions of SRs that the SRA must satisfy.
   • Domain Concepts (DC): Core entities and relationships (e.g., data sources, processing layers).
3. Architectural Synthesis: Defining the SRA through architectural views (design-time modules, runtime components, deployment) using diagrams (e.g., UML) and textual specifications.
4. Architectural Evaluation: Validating the SRA via traceability mapping, surveys on perceived usefulness, and tool refactoring experiments.

Specific Contribution of This Paper:
The authors executed Step 1 (Information Source Investigation) specifically focusing on Stakeholders. They conducted a focus group session at the AIRE'25 workshop to elicit initial requirements.

• Participants: 20 researchers and practitioners.
• Process: Two iterations of group discussions using a 2D matrix (Priority vs. Complexity) to rank requirements.
• Categories Analyzed: Functional requirements, non-functional requirements, domain concepts, and challenges related to Generative AI.

4. Key Results

The focus group activity yielded 36 key system-level requirements (SRs) that validate the need for an SRA. These requirements are categorized by the four improvement dimensions of the authors' vision:

• For Improved Reuse:
  • Need for modularity and component reuse (SR11).
  • Requirement for black-box interfaces to hide implementation details (SR07).
  • Standardized reusable input parsers for common formats (SR02, SR38).
  • Transparent intermediate processing steps (SR06).
• For Improved Applicability:
  • Configurability and adaptability to specific RE artifacts (SR03).
  • Multilingual support (SR04, SR05).
  • Explicit communication of input limitations (SR25).
• For Improved Comparability:
  • Support for evaluation and benchmarking (SR01).
  • Reproducible behavior under fixed conditions (SR17).
  • Version tracking (SR18) and durable storage of artifacts/results (SR15).
• For Improved Maintenance:
  • Robustness and recoverability (SR14, SR15).
  • Compliance with legal and data protection standards (SR08, SR09).
  • Long-term version control strategies (SR18).

5. Significance and Future Work

• Theoretical Contribution: This paper is the first to explicitly propose and outline a roadmap for a Software Reference Architecture (SRA) specifically for the NLP4RE domain. It moves the field from ad-hoc script development toward standardized engineering.
• Practical Impact: By defining a shared blueprint, the SRA will reduce development effort, lower the barrier to entry for new researchers, and enable rigorous scientific comparison of NLP techniques in RE.
• Open Science: All materials, including the raw focus group data and consolidated requirements, are publicly available via GitHub and Zenodo to ensure reproducibility.
• Next Steps: The authors plan to derive specific Architectural Requirements (ARs) from the elicited system requirements, assess existing tools against this new framework, and iteratively design and implement the SRA in collaboration with the community.

In summary, this paper establishes the foundational evidence and strategic roadmap necessary to transform NLP4RE tooling from fragile, isolated academic projects into a sustainable, modular, and interoperable software ecosystem.