Speaker
Description
This paper presents a case study exploring the potential of Large Language Models for learner-oriented phraseography through the generation of COBUILD-style Full-Sentence Definitions and AI-generated dual illustrations representing both the literal and phraseological meanings of idioms. The study is situated within Phrasolino, an electronic Italian-German phraseological learner’s dictionary for German-speaking children aged 7–11 learning Italian as an L2. A corpus-based dataset of 100 Italian idioms containing the body-part noun testa (‘head’) was compiled from the Italian Web Corpus itTenTen20 and complementary lexicographic resources. ChatGPT-5.5 was prompted to generate child-oriented Full-Sentence Definitions and corresponding dual illustrations. The outputs were evaluated qualitatively with regard to semantic accuracy, phraseological adequacy and pedagogical suitability. The findings show that ChatGPT successfully reproduces the formal conventions of COBUILD-style definitions, although recurrent difficulties emerge in the semantic simplification and age-appropriate adaptation of phraseological meaning. The image-generation analysis further reveals both the potential and the limitations of AI-assisted multimodal phraseography, particularly regarding semantic disambiguation, literal interpretation and representational bias. Overall, the study highlights the opportunities of generative AI for learner-oriented lexicography while emphasizing the continuing necessity of lexicographic supervision and post-editing.