Speaker
Description
The past few years have seen numerous attempts to make use of generative artificial intelligence (AI) in the field of lexicography. One such field is neology – the detection of new words and the preparation of their lexicographic entries for further study (cf. Zheng et al., 2024). When identifying neologisms in large collections of texts with the help of AI, existing neologism data – usage examples already collected by lexicographers – can be used for two purposes: (1) as positive examples included in AI prompts and (2) as material for evaluating model performance (cf. Tang et al., 2025). Since the number of neologisms in a dataset of usage examples is known, it is possible to calculate not only precision but also recall, enabling an objective evaluation of model performance.
Traditional methods of collecting neologisms are still very effective. The Database of Lithuanian Neologisms (DLN) relies on two traditional collection methods: editors regularly record neologisms encountered while reading news portals and social media; members of the general public submit neologism candidates they have identified via a form on the DLN website. As of 2026, the DLN contains more than 12,000 detailed neologism entries with definitions and usage examples. Because individual entries may have several examples, the database currently includes more than 43,000 usage examples in total. Many of these examples consist of multiple sentences, thus making the DLN a source for a very neologism-rich corpus comprising several million words. An additional advantage of manual collection is that neologisms can be documented from sources that are difficult to access for automated tools: social media posts, comments on YouTube and news portals, radio and TV recordings.
A dataset of this size and quality provides a strong basis for developing AI tools for automated neologism extraction from web texts and definition generation.
This presentation discusses prompt engineering experiments aimed at identifying an optimal prompt for the task of neologism detection, incorporating usage examples from the DLN. The experiments focus on the component of the prompt known as the instruction (or task description). Prompt engineering guidelines (Liu et al., 2024) typically describe a prompt as consisting of several elements: (1) the instruction, which tells the model what task to perform; (2) a role specification defining the model’s assumed role; (3) examples with corresponding correct responses; (4) formatting, style and length requirements for the output; (5) the input material that the model is asked to process.
Two alternative instruction strategies for neologism detection were tested. The first strategy instructs the model to process a given text and return only the neologisms it identifies (extraction-style). The second strategy instructs the model to examine each word in the text individually and determine whether it constitutes a neologism (classification-style). Using an open-weight language model, a series of experiments was conducted to evaluate the performance of both prompting strategies. Although the difference between these formulations may appear relatively minor, it results in substantial differences both in output quality and in the computational resources required to complete the task.
The experiments demonstrate that both prompting strategies achieve satisfactory performance in neologism detection. The classification-style prompt shows a slight advantage in the total number of correctly identified neologisms, whereas the extraction-style prompt demonstrates a significant advantage in terms of lower noise, i.e. fewer words incorrectly identified as neologisms. An even greater difference between the two strategies emerges from the perspective of practical application. When processing the same text on identical hardware, the extraction-style prompt required nearly twenty times less processing time than the classification-style prompt.
Both prompting strategies exhibit similar weaknesses. Despite explicit instructions, the model may return words containing introduced misspellings, provide grammatically altered forms instead of the original forms occurring in the analyzed text, omit relevant words from the output, reorder words relative to their order in the source text, or introduce words that do not appear in the analyzed text at all (hallucinations).
Further research should include a lexicographic evaluation of how the two prompting strategies perform in detecting specific categories of neologisms, such as derivatives, compounds, and borrowings. From a technical perspective, methods for mitigating model-induced distortions of analyzed words should also be explored.
The experiments indicate that current large language models can provide substantial assistance to lexicographers in the collection of neologisms. Nevertheless, professional linguistic expertise remains necessary for the final evaluation of candidate neologisms and the removal of false positives.