Building another Spanish dictionary, this time with GPT-4

The paper presents Spanish Built Factual Freecianary 2.0 (Spani-BFF-2), an updated Spanish dictionary generated using GPT-4-turbo, which improves upon its predecessor by addressing limitations, expanding coverage, and incorporating part-of-speech tags. The study evaluates GPT-4-turbo's performance in generating definitions, comparing it to the Diccionario de la Lengua Española (DLE) and analyzing its strengths and weaknesses, particularly in handling monosemy and polysemy. While GPT-4-turbo generally provides accurate definitions, it struggles with polysemy and occasionally hallucinates. The dictionary has a higher precision for monosemy but lower recall for polysemous words, indicating room for improvement in capturing multiple meanings. The study also identifies issues with subword tokenization and error analysis, suggesting future work on refining the model's handling of rare and unconventional words, as well as responsible use of large language models in lexicography.