Linguistic Evaluation of Large Language Models: Insights and Error Typologies from Polish Matura Exam Responses

LINGUISTIC EVALUATION OF LARGE LANGUAGE MODELS: INSIGHTS AND ERROR TYPOLOGIES FROM POLISH MATURA EXAM RESPONSESThe article presents a manual analysis of texts generated by four large language models (LLMs): Bielik 2.3, GPT-4, LLaMA 3.1, and PLLuM 8x7B. The texts are in Polish and the text topic is based on the Polish Matura exam. This research extends previous work by Rafał Mazur (2024), who tested only one LLM (Chat GPT 3.0) and had not formulated his own typology of errors. In the article, a typology of language errors formulated for the purposes of this study is used, based on existing solutions (Buttler 1971; Markowski 2007, 2014; Burkacka 2014; Pędzich 2014a, 2014b), as well as quantitative and qualitative criteria concerning the model’s references to readings and cultural contexts. The study found that although GPT-4 made the fewest linguistic errors and did not hallucinate, it referred to the fewest number of texts, including no Polish readings or Polish cultural contexts. The PLLuM 8x7B, on the other hand, referred to the largest number of Polish readings of all models tested.