Year: 2024
Type: conference-paper
Abstract: The study deals with token classification, focusing on BERT-based models, in particular multilingual BERT (mBERT) and BERTić. We take pre-trained models and fine-tune them for a specific downstream task, namely the recognition of metaphorical collocations. Metaphorical collocations are a type of lexical collocation where the base word retains its literal meaning, while the collocate takes on a figurative meaning. The hrWaC was used to search a web-based Croatian corpus and create collocational profiles of the selected most frequent nouns. The manually annotated lists were used as the basis for creating the training and test sets. The resulting fine-tuned models achieved promising results, and error analysis highlighted the task's inherent difficulty. In the next phase of the research, we will apply the fine-tuned model to other languages but also retrain the model on the higher quality dataset obtained by conducting a second annotation pass.
Institution: University of Rijeka
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
Funder: Hrvatska Zaklada za Znanost
OpenAlex ID: W4409403743