Logo image
Back
SmolLab_SEU at CheckThat! 2025: How Well Do Multilingual Transformers Transfer Across News Domains for Cross-lingual Subjectivity Detection?
Conference proceeding   Open access

SmolLab_SEU at CheckThat! 2025: How Well Do Multilingual Transformers Transfer Across News Domains for Cross-lingual Subjectivity Detection?

Md. Abdur Rahman, Md. Al Amin, Md. Sabbir Dewan, Md. Jahid Hasan and Md Ashiqur Rahman
CEUR Workshop Proceedings, Vol.4038
Working Notes of the Conference and Labs of the Evaluation Forum (CLEF 2025) (Madrid, Spain, 09/09/2025–12/09/2025)
09/2025
pdf
Published1.36 MBDownloadView
CC BY V4.0,  Open Access

Abstract

Automated detection of subjectivity in news articles is an important problem for fighting against fake news and promoting journalistic accountability, but this is a challenging task in various linguistic settings. This paper shows our method on Task 1: Subjectivity of the CLEF 2025 CheckThat! Lab to identify objective information vs subjective opinion in news content available in several languages. To this aim, we experimented with a variety of Transformer models using architecture specifically designed for a particular language (GermanELECTRA-large, MARBERT-v2, RoBERTa-large) as well as multilingual (XLM-RoBERTa-large, mDeBERTaV3-base, InfoXLM-large) and zero-shot (mBERT-base) models. Our approach utilized the fine-tuning of pre-trained models with hyperparameters and class-weighted loss functions so as to tackle the imbalanced data. Experimental results show that our models perform well: German-ELECTRA-large achieved 0.8520 F1 in German, XLM-RoBERTa-large got 0.8356 F1 in Italian and 0.8040 in Romanian zero-shot, RoBERTa-large is best, with 0.7948 F1 on English, and InfoXLM-large achieves 0.7114 F1 on multilingual setting. In official ranking, our systems obtained 1st rank in Monolingual German, 2nd in Zero-shot Romanian, 3rd in Monolingual Italian, 5th in Zero-shot Ukrainian, 6th in Multilingual, 8th in Zero-shot Polish and 9th in Monolingual Arabic and English. Error analysis shows that monolingual models excelled monolingually, and multilingual architectures achieve better cross-lingual generalization in the zero-shot settings. This work provides insights into the suitability of Transformer in multilingual subjectivity detection and demonstrates the difficulties in recognizing subtle subjective cues in different linguistic environments. © 2025 Copyright for this paper by its authors.

Details

Metrics

1 File views/ downloads
9 Record Views
Logo image