Category-Dependent Effectiveness of Web-Based and AI-Generated Usage Examples in Modern Greek Pedagogical Lexicography

Background. Usage examples are an important feature of pedagogical dictionaries, as they help learners understand meaning, context, and lexical use. Although web-based resources and large language models offer new possibilities for generating dictionary examples, their pedagogical suitability has not yet been sufficiently examined in Modern Greek. Purpose. To address this gap, this exploratory study compares Web-based and AI-generated usage examples across three lexically challenging categories in Modern Greek: polysemous verbs and nouns, and idiomatic expressions. It examines whether and how the pedagogical suitability of each source varies according to lexical complexity for learners aged 11-15 and explores their potential integration in pedagogical lexicography. Materials and Methods. A small-scale exploratory dataset of 81 usage examples was collected from Web sources and generated by GPT-5.3 Instant and DeepSeek-V3 and was evaluated using a combined framework drawing on GDEX criteria and pedagogical appropriateness indicators on a three-point ordinal scale. A learner validation phase with 41 students (aged 11-15) was also conducted. Results. Within the analysed dataset, the findings suggest category-related differences across lexical categories. Web-derived examples showed greater semantic diversity and higher correct meaning identification rates in the learner evaluation (56.50%), particularly for figurative and idiomatic uses. AI-generated examples, especially those produced by GPT-5.3, showed higher intelligibility scores and performed more consistently for semantically stable items such as polysemous nouns, but exhibited limitations in idiomatic language, with correct meaning identification for GPT-5.3 dropping to 32.37% for specific idiomatic items. DeepSeek-generated examples showed the most pronounced limitations within the analysed categories, particularly in typicality and pedagogical appropriateness. Conclusion. The findings suggest that, within the analysed dataset, the pedagogical value of usage examples depends not only on their source, but also on the type of lexical complexity involved. Neither Web-based nor AI-generated examples can independently meet the requirements of a learner-oriented dictionary. Instead, the findings support a hybrid category-sensitive approach that combines Web-based authenticity, expert curation, selective AI assistance, and attention to learners’ actual comprehension. The methodological framework developed here may offer a basis for future comparative work in other morphologically complex and idiomatically rich languages, though transferability remains to be tested.