Speakers
Description
Digital dictionaries can become more than consultation tools when their internal data are sufficiently structured to be reused and transformed into embedded learning components. This paper evaluates whether large language models (LLMs) can convert PORTLEX entries, represented as reviewed JSON files, into interactive HTML didactic modules integrable into their corresponding dictionary entries. PORTLEX is a multilingual valency-oriented dictionary of the nominal phrase, and its microstructure offers a controlled source for activities on formal links, ordering, complement function and semantic compatibility. The experiment compares a GPT-based custom system and a Gemini-based Gem across three languages, four semantically aligned nouns per language and two model outputs per entry, producing 24 modules and 48 expert evaluations. Results show that both systems generated broadly usable modules and generally preserved the linguistic-valency information encoded in the JSON input. The main limitations did not concern basic HTML functionality, but pedagogical regulation: feedback, correction, scaffolding, distractor quality and ambiguity management. Overall scores did not differ significantly; only feedback and correction favoured GPT after statistical adjustment. Language-based patterns were diagnostic rather than conclusive. The results support treating LLM-generated modules as controlled prototypes within a documented, reusable workflow, not as ready-made pedagogical products.