AI Wizards at CheckThat! 2025: Enhancing Transformer-Based Embeddings with Sentiment for Subjectivity Detection in News Articles
Matteo Fasulo, Luca Babboni, Luca Tedeschini
2025-07-17
Summary
This paper talks about enhancing transformer-based classifiers by adding sentiment information to better detect subjectivity in news articles, especially across multiple languages.
What's the problem?
The problem is that detecting whether a sentence is subjective (expressing opinions or feelings) or objective (stating facts) is hard because models struggle to understand subtle language differences and work well in many languages without specific data.
What's the solution?
The authors improved transformer models by including sentiment scores from an extra model as part of the input, which helps the classifier better recognize subjective content. They tested this method in different languages and scenarios, including where the model hasn't seen the language before, achieving strong performance.
Why it matters?
This matters because accurately spotting subjective sentences helps with fact-checking and combating misinformation in news, making information more trustworthy and helping people better understand what is opinion versus fact.
Abstract
Sentiment-augmented transformer-based classifiers improve subjectivity detection in multilingual and zero-shot settings, achieving high performance and ranking first for Greek.