Neural Metaphor Detection for Slovene
Keywords:Token-level metaphor detection, Slovene, Transformers, Multilingual detection, Cross-lingual detection
AbstractMetaphors are linguistic expressions using comparison with another concept to potentially improve the language expressivity. Due to relevant downstream applications, metaphor detection is an active topic of research. Most of the research is focused on English, while other languages are less covered. In our work, we focus on Slovene, presenting the first word-level metaphor detection experiments. We apply multiple transformer-based large language models on four versions of two publicly available Slovene corpora: KOMET and G-KOMET. We perform monolingual, multilingual, and cross-lingual experiments, using the VU Amsterdam metaphor corpus as an additional source of metaphor knowledge. We evaluate the models quantitatively using word-level $F_1$ score and find that (1) the most consistently well-performed model is the trilingual CroSloEngual BERT model, (2) the addition of English data in multilingual experiments does not improve the performance significantly, and (3) the cross-lingual models achieve significantly worse results than their monolingual and multilingual counterparts.
Copyright (c) 2023 Matej Klemen, Marko Robnik-Šikonja
This work is licensed under a Creative Commons Attribution 4.0 International License.