Norma, ús i interferència: biaixos lingüístics en els models de llenguatge en català

Almena Rodríguez, Mireia

Norma, ús i interferència: biaixos lingüístics en els models de llenguatge en català

License

Citation

Almena Rodríguez, M. Norma, ús i interferència: biaixos lingüístics en els models de llenguatge en català. 2025. handle: http://hdl.handle.net/10230/70986

This citation was generated automatically.

Abstract

This study analyzed four large language models (LLMs) to assess their linguistic quality in Catalan and to examine potential interference from Spanish. It compares one monolingual model (Catalan BERTa-v2) with three multilingual models (X-MOD, TwHIN-BERT and XLM-V). The analysis has been carried out through a Python program, which shows that the monolingual model aligns most closely with the Catalan norm, although it also reflects non-normative expressions used frequently by speakers, also suggesting a descriptive use of the language. Additionally, sentences from the CONTACT project, which focuses on Catalan as a heritage language in bilingual contexts, have been used in order to analyze if the models are capable of detecting linguistic contact patterns. The findings reveal a certain ability to identify grammatical structures in anaphorical contexts, but no consistent patterns were found in the case of differential object marking or the use of verbs ser and estar. Overall, the study demonstrates that LLMs generate both normative and non-normative structures, which raises important questions about their role in the preservation and normalization of Catalan, as well as the need to regulate their development to ensure an adequate linguistic use.