| Abstract | Stance classification can be used in various scenarios, such as fake news detection or public opinion measurement. However, little work has been done on stance detection in multilingual data. For this reason, this work uses a multilingual, multi-target, and multi-topic dataset to develop a classifier for detecting stance in such data. The classifier was trained using pre-trained BERT models, with various experiments showing superior performance of a fine-tuned multilingual BERT model with self-training. Since the dataset was unbalanced, with the main label being "in favor", the macro-averaged F1 score was used for measurement. The best performing model achieved a macro-average F1 score of 0.8862 using the same proposals in stance classification for training and testing. The same approach was used to train two classifiers for the CLEF 2023 Touché Lab Task 4 Subtask 1 and 2, using new, previously unseen proposals for testing. However, by using new unseen proposals, the results deteriorated significantly, and in the challenge only a macro F1 score of 0.324 and 0.417 was achieved. |
|---|
