DysLexLens: A Low-Resource LLM Framework for Analysing Dyslexic Learners Insights from Online Forums

Synthesis: DysLexLens is a low-resource Large Language Models (LLMs) framework designed to analyze how dyslexic learners experience AI tools by mining online forum discussions. The framework employs dictionary-driven filtering to construct focused corpora from Reddit, integrates LLM-assisted knowledge graph reasoning, and generates verifiable query responses about learners' lived experiences with AI for reading, writing, and study tasks.

The research reveals that while dyslexic learners find value in AI tools for supporting literacy, they face significant accessibility barriers including inconsistent output quality and lack of equitable accommodations. This has direct implications for student-AI interaction design and suggests that inclusive AI education must address language and literacy support across both K-12 and higher education settings.

By grounding analysis in real user discourse rather than controlled experiments, the work complements traditional student modeling approaches and provides an evidence base for designing AI tools that better serve neurodiverse learners.

What this means for practice

• Learners. Treat every AI answer as a draft to check against a source rather than a finished product: the audited responses averaged 0.75 answer relevancy but only 0.43 response groundedness, and 6 of 100 audited claims could not be traced to any evidence.
• Rewrite a query with the exact terms of your own material when the output looks off: keyword-perturbed queries dropped to 0.34 mean answer relevancy, while paraphrased queries held at 0.58, so substituting vocabulary matters more than rephrasing the same words.
• Ask for the passage behind each claim and check it yourself, because only 10 of 11 main responses were fully verifiable while 56 of 89 follow-up responses were only partially verifiable.
• Log which tools help with which literacy tasks and which fail, then bring that record to disability services or instructors, so accommodation choices rest on your own reported experience rather than vendor claims.
• Do not read a reliable-sounding answer as an accessible one: the study finds learners describe existing AI tools as useful but still limited, so push back on tools that produce inconsistent output quality or no accessibility options.

Limitations

• The evidence base is 319 filtered posts from 27 subreddits, cut down from an initial corpus of 23,480 posts and comments (1,663,250 words) across 45 subreddit communities; the authors state that Reddit discussions do not represent all dyslexic learners.
• Retrieval precision and claim-level grounding remained the pipeline's main weaknesses: across 30 responses the mean Context Relevancy was 0.40 and Response Groundedness 0.43, RQ3 scored zero Response Groundedness, and RQ5 scored zero Context Relevancy.
• The human audit of 100 claims found 39 fully verifiable, 55 partially verifiable and 6 not verifiable, with follow-up rows far weaker than main responses (56 of 89 only partially verifiable, mostly because rows exported the full retrieved chunk instead of a short exact evidence phrase).
• Both knowledge-graph construction and every generated response used gpt-4o-mini, and the authors note that LLM-based triple extraction and retrieval may introduce noise, so the graph should be treated as an interpretive aid rather than a complete representation of learner experience.

Citation

Dana Rezazadegan, Atie Kia, Phongpadid Nandavong, Dominique Carlon, Jeremy Nguyen (2026). DysLexLens: A Low-Resource LLM Framework for Analysing Dyslexic Learners Insights from Online Forums. Artificial Intelligence (cs.AI).