Assessing the Impact of CLEAR Prompt Engineering on LLM-based Term Extraction in Welsh and English

This study explores the effectiveness of the CLEAR (Concise, Logical, Explicit, Adaptive, Reflective) prompt engineering framework [1, 2] for extracting technical terms from Welsh texts using ChatGPT-5. In the contemporary digital landscape, language technologies for minoritised languages like Welsh continue to face challenges, particularly in the development of high-quality terminological resources and automated term extraction methods.
Using an experimental methodology, the study compares ChatGPT-5 outputs across four different prompting conditions: unstructured Welsh and English prompts, and structured prompts in both languages following the complete CLEAR framework. Extracted terms are evaluated against the Termiadur Addysg, a Welsh-language dictionary of educational terminology, focusing on the legal domain.
Results indicate that the use of the CLEAR framework is associated with a statistically significant increase in the proportion of terms that match entries in the Termiadur Addysg. No statistically significant difference was observed between Welsh and English prompt outputs. Qualitative analysis further suggests that CLEAR-based prompts are also associated with fewer erroneous, misspelled, and ambiguous terms than unstructured prompts.