Speakers
Description
In recent years, several lexicographic studies have addressed the question of which lexicographic processes can effectively make use of LLMs (De Schryver, 2023; Lew, 2024; McKean & Fitzgerald, 2024). Some studies are dedicated specifically to Slovenian lexicography (Gantar, 2024; Arhar Holdt et al., 2025), including Meterc & Jakop (2025) who have examined the use of LLMs in the compilation of the Dictionary of the Slovenian Standard Language eSSKJ (2016–), which is being compiled at the Fran Ramovš Institute of the Slovenian Language, ZRC SAZU, and published on the Fran dictionary portal (Ahačič et al., 2015).
This paper focuses on the performance of ChatGPT 5.5 in compiling eSSKJ dictionary entries for classifying adjectives. According to the editorial principles used in eSSKJ, classifying adjectives tend to have relatively uniform type-based definitions and predictable semantic groups of collocates, which makes them a suitable case study for testing the extent to which LLMs can successfully generate complete dictionary entries. The study tests the LLM’s performance in the compilation of 35 dictionary entries for classifying adjectives derived from nouns denoting sports, academic disciplines, and animals (e.g. košarkarski ‘pertaining to basketball’, geološki ‘pertaining to geology’, kengurujev ‘pertaining to kangaroo’), selected because the corresponding base nouns are often monosemous or exhibit regular polysemy.
The LLM was provided with the following background knowledge prior to prompting: (1) a typology of definitions used for classifying adjectives, (2) a training set of 72 previously published eSSKJ dictionary entries for classifying adjectives, (3) corpus data in the form of word sketches for these adjectives, and (4) up to 1,000 randomly selected concordances for each adjective from the Gigafida 2.0 Slovenian reference corpus (where fewer concordances were available, all examples were provided). Apart from the typology of definitions, which was supplied to the LLM in textual form, all other data were encoded in XML. The same types of data had originally been available to human editors during the compilation of the dictionary entries.
We evaluated the performance of the LLM in the following editorial tasks: (1) sense differentiation, (2) formulation of (type-based) dictionary definitions, (3) selection of typical collocations and their assignment to the appropriate dictionary senses, and (4) illustration of senses with typical dictionary examples drawn from the Gigafida 2.0 corpus. As a control measure, the same dictionary entries were also edited manually in parallel and were not uploaded to the LLM.
Testing the performance of the LLM in its editorial treatment of classifying adjectives in eSSKJ shows that it formulates type-based dictionary definitions in accordance with the established typology, although the ordering of predictable metonymic senses does not always follow the established hierarchy. It also selects appropriate collocates, both in terms of the range of collocates included and their assignment to dictionary senses, and successfully adheres to formatting guidelines for collocations (semantic clusters, formal and alphabetical ordering, etc.). The model also performs well in the selection of dictionary examples. In these respects, the LLM’s editorial treatment is comparable to dictionary entries produced by human editors. The LLM is also successful in distinguishing senses in dictionary entries that contain only predictable, type-based senses; however, it often fails to identify additional metaphorical or derived senses, as well as fixed expressions.
The results suggest that although ChatGPT 5.5 does not yet perform all tasks involved in compiling entries for classifying adjectives in eSSKJ entirely satisfactorily, it nonetheless represents a highly useful tool for lexicographers. At present, it can assist eSSKJ editors particularly with more routine tasks, which is especially valuable because these tend to be very time-consuming.