Developing a methodology for verifying frame structures based on expert evaluations by large language models

The purpose of the study is to develop a methodology for verifying frame structures based on data from large language models. The article proposes a methodology for using generative artificial intelligence to verify a candidate frame structure constructed through frame modeling based on the analysis of dictionary definitions. The authors focus on the procedure and criteria for comparing the candidate frame structure with the structure obtained through interaction with generative artificial intelligence, as well as on the principles of frame modeling based on large language model data. The scientific novelty lies in proposing, for the first time, a procedure for verifying a frame model constructed using the classical method of dictionary-definition analysis. The procedure involves generating alternative frame models by several multilingual models using a single detailed prompt, followed by mandatory corpus-based verification of the proposed components. The study establishes that, when the candidate frame model “scientific and technological achievement” is compared with structures generated by five language models (Alice AI, GigaChat, ChatGPT, Perplexity AI, and DeepSeek), semantically obligatory and lexicographically explicit components are consistently reproduced regardless of the language of the model’s training corpus, whereas discourse-pragmatic components remain beyond the reach of generative inference. Of the five new components proposed by the models, only one passed empirical verification: the “Visual and Narrative Elements” slot, which was incorporated into the verified frame model.