A new study on polysemy in language models stands out in the international reference scientific forum
The predoctoral researcher Anna Temerko defends in Vienna a new contribution from CiTIUS focused on how AI models process regular polysemy motivated by metaphors during the CoNLL 2025 conference, the world's most prestigious event in the field of artificial intelligence applied to language.
Anna Temerko, a predoctoral researcher at CiTIUS, recently participated as a speaker at the international conference CoNLL 2025, held in Vienna (Austria) as part of the ACL (Association for Computational Linguistics) conference, considered the most prestigious global event in the field of Natural Language Processing.
The presented paper, A Continuous Approach to Metaphorically Motivated Regular Polysemy in Language Models, is already listed as one of the published works in the conference proceedings. The study analyzes how the most advanced artificial intelligence models—like ChatGPT, Gemini, or Llama— manage regular polysemy, a phenomenon where the same word can systematically acquire different meanings. In this case, the research focuses on metaphorically motivated polysemy, that is, when a new meaning arises from a metaphorical extension (for example, when the verb illuminate is used both to “give light” and “bring clarity to an idea”).
The results show that these models are able to reflect this regularity in a manner very similar to human behavior: they distinguish between types of senses, recognize different degrees of regularity, and reliably reproduce the assessments that people would make.
The work was developed in collaboration with CiTIUS researchers Marcos García, a specialist in corpus linguistics and linguistic variation, and Pablo Gamallo, an international reference in computational semantics and the development of linguistic resources.
This new contribution reinforces the international projection of CiTIUS, a center co-financed by the European Union through the Galicia Feder Program 2021-2027, and consolidates its role as a benchmark in the study of language from technical and cognitive perspectives.
Anna Temerko, a predoctoral researcher at CiTIUS, recently participated as a speaker at the international conference CoNLL 2025, held in Vienna (Austria) as part of the ACL (Association for Computational Linguistics) conference, considered the most prestigious global event in the field of Natural Language Processing.
The presented paper, A Continuous Approach to Metaphorically Motivated Regular Polysemy in Language Models, is already listed as one of the published works in the conference proceedings. The study analyzes how the most advanced artificial intelligence models—like ChatGPT, Gemini, or Llama— manage regular polysemy, a phenomenon where the same word can systematically acquire different meanings. In this case, the research focuses on metaphorically motivated polysemy, that is, when a new meaning arises from a metaphorical extension (for example, when the verb illuminate is used both to “give light” and “bring clarity to an idea”).
The results show that these models are able to reflect this regularity in a manner very similar to human behavior: they distinguish between types of senses, recognize different degrees of regularity, and reliably reproduce the assessments that people would make.
The work was developed in collaboration with CiTIUS researchers Marcos García, a specialist in corpus linguistics and linguistic variation, and Pablo Gamallo, an international reference in computational semantics and the development of linguistic resources.
This new contribution reinforces the international projection of CiTIUS, a center co-financed by the European Union through the Galicia Feder Program 2021-2027, and consolidates its role as a benchmark in the study of language from technical and cognitive perspectives.