Event extraction is a natural language processing task aimed at systemati
cally identifying structured information about events from unstructured text,
including triggers, participants, temporal context, and location. Most existing
approaches and annotated datasets are limited to English, hindering the
development and evaluation of models for other languages. In this thesis, we
address event extraction in Slovenian using two approaches. We first evaluate
zero-shot prompting across several publicly available large language models
of varying sizes and providers. We then fine-tune the models using Low-Rank
Adaptation (LoRA) and compare performance before and after adaptation.
The WIKIEVENTS, CASIE, and ROBUST datasets are translated into
Slovenian using the NLLB-200 machine translation model, and inference is
performed on the translated versions. InstructUIE and ADELIE-DPO serve
as reference models. The results show that large language models, despite
fine-tuning, fall short of models specialized for event extraction. The latter
achieve higher results on datasets in their native language, however they
exhibit a substantially larger drop in performance when applied to Slovenian.
|
