Application of small language models for semantic analysis of Web interface accessibility

Abstract

Web accessibility remains a critical aspect of ensuring equal opportunities for internet resource usage, especially for people with disabilities. The Web Content Accessibility Guidelines 2.5.3 criterion “Label in Name” requires that the accessible name of an interface component include text that is visually presented. Existing automated verification methods for this criterion are predominantly based on primitive string comparison, which does not account for semantic context. Objective: investigate the possibilities of using small language models with up to 1 billion parameters for automated semantic analysis of compliance with Web Content Accessibility Guidelines 2.5.3, as an alternative to resource-intensive large language models and limited algorithmic methods. Methodol ogy: the research involved creating synthetic datasets (7,200 English-language and 5,615 Ukrainian-language samples) and using real-world datasets (Top500 – 380 samples, UaUniv – 319). Sentence Bidirectional Encoder Representations from Transformers models were tested for computing semantic similarity, and fine-tuning of the google/electra-base-discriminator model was performed for 3-class classification of semantic relationships (“similar”, “unrelated”, “opposite”). Results: the trained model of 437 MB demonstrated high accuracy on syn thetic data (0.96) and sufficient accuracy on real datasets (Top500: 0.77, UaUniv: 0.73). The model effectively identifies all three classes of semantic relationships with an accuracy of 95.1 % for “opposite”, 92.7 % for “un related”, and 97.4 % for “similar” texts in the validation sample. Conclusions: the research confirmed the feasi bility of using small language models for automated verification of semantic compliance according to Web Content Accessibility Guidelines 2.5.3. The proposed approach provides acceptable classification accuracy with significantly lower computational costs compared to large language models, allowing for the integration of se mantic analysis into standard development and testing processes. Despite certain limitations, the developed so lution can significantly improve web accessibility testing.

Prombles in programming 2025; 2: 77-86