Intelligent Approaches in the Software Development Process: A Systematic Literature Mapping
Luciano Arruda Teran, Alan Trindade de Almeida Silva, Giselle Lorrane Nobre Melo, Marcelle Pereira Mota · 2021 · Proceedings of the X Latin American Conference on Human Computer Interaction (CLIHC) · doi:10.1145/3488392.3488397
Summary
This short paper from the Federal University of Pará in Brazil presents a systematic literature mapping (SLM) investigating how intelligent approaches — methods and tools using artificial intelligence concepts — are used to assess and validate accessibility requirements during the software development process. The authors searched ACM Digital Library and IEEE Xplorer using terms combining software/web/mobile development with accessibility and testing/evaluation/validation, collecting 2,627 initial results. After applying inclusion and exclusion criteria (abstracts must contain search keywords, primary studies only, English or Brazilian Portuguese, software engineering or HCI thematic area), 16 studies were selected for analysis. The mapping was structured around five research questions addressing how approaches assist developers in validating accessibility requirements, whether they improve access for people with disabilities, which accessibility guidelines are used, whether the approaches are considered intelligent, and what AI concepts are employed.
Key findings
The most striking finding was the near-complete absence of intelligent (AI-based) approaches to accessibility validation in the software development process. Of the 16 selected studies, only 2 used methods that could be considered intelligent: one employing computer vision and machine learning, and another using supervised learning with a decision tree tool called "CrowdRex." The remaining studies relied on traditional methods for validating accessibility requirements. WCAG was the most cited accessibility guideline (6 of 16 studies), while Axe 3.4 and WAI-ARIA were each cited once. However, half of the selected studies (8 of 16) did not mention which accessibility guidelines they used. The W4A conference contributed the most papers (4 of 16). Geographically, the USA led with 31.25% of publications (5 studies), followed by Brazil, UK, Singapore, and Switzerland with 12.5% each (2 studies each). Publication years peaked in 2015 and 2019 (3 studies each). Among the traditional approaches found, continuous and proactive testing integrated throughout development sprints was highlighted as effective, with one study noting that validating accessibility at each development sprint generates more satisfaction for people with disabilities than only addressing issues when reported. Several studies emphasized that automated tools should handle routine checks, freeing accessibility specialists for more advanced manual analysis.
Relevance
This mapping study exposes a significant gap in the intersection of AI and accessibility: despite the rapid advancement of machine learning and AI techniques, these capabilities are barely being applied to automate accessibility validation during software development. This is notable because accessibility testing remains heavily manual and expert-dependent, creating a bottleneck that intelligent automation could help address. The finding that half of the reviewed studies did not reference any specific accessibility guideline suggests a disconnect between accessibility research and standards compliance in practice. For practitioners and researchers, this paper serves as a call to action: there is substantial untapped potential for applying computer vision (to evaluate visual accessibility), natural language processing (to assess content readability and alt text quality), and machine learning (to predict accessibility issues from code patterns) to the software development lifecycle. The emphasis on integrating accessibility validation into agile development sprints — rather than treating it as an afterthought — aligns with shift-left testing principles that are increasingly important in modern software development. The study's limitations include a small corpus (16 papers from only 2 databases) and the rapid pace of AI development since 2021, meaning the landscape may have already shifted significantly.
Tags: software engineering · accessibility testing · artificial intelligence · automated testing · systematic review · software development · WCAG compliance
Standards referenced: WCAG · WAI-ARIA · Axe 3.4