Multiclass Relation Extraction with Deep Learning Models
Sharma, Tanvi (2026)
Tampere University
2026
Tieto- ja sähkötekniikan tohtoriohjelma - Doctoral Programme in Computing and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Väitöspäivä
2026-10-07
Julkaisun pysyvä osoite on
https://urn.fi/URN:ISBN:978-952-03-4774-1
https://urn.fi/URN:ISBN:978-952-03-4774-1
Tiivistelmä
The rapid spread of misinformation across digital platforms presents major societal challenges, particularly in domains such as healthcare, politics, and finance. While conventional misinformation detection often relies on binary classification, this thesis addresses misinformation quantification through multiclass relation extraction, enabling a more fine-grained and interpretable analysis of relationships between entities in text.
This thesis makes four primary contributions. First, it provides a comprehensive methodological review of relation extraction, identifying key trends, challenges, and research gaps. Second, it introduces manually curated COVID-19 multiclass relation extraction datasets from Reddit and PubMed, enabling systematic evaluation of transformer-based models, with BioRedditBERT and BioBERT achieving the highest F-scores on the Reddit and PubMed datasets, respectively. Third, the thesis presents COVID-MisinfoRel, a novel manually annotated misinformation dataset that frames misinformation as a multiclass relation extraction problem and demonstrates that transformer models outperform conventional machine learning approaches, such as Random Forest and Support Vector Machine, in distinguishing misinformation categories. Finally, this thesis extends misinformation quantification across multiple domains, including crime, entertainment, finance, and politics, by integrating semi-supervised learning and topic modeling to reduce annotation effort and refine misinformation categories. Furthermore, it presents the first application of the Mamba state-space architecture to quantify misinformation for multiclass relation extraction and also proposes a novel Hybrid Transformer–Mamba model that combines contextual representation learning with efficient sequential modeling to improve scalability and cross-domain generalization.
Overall, this thesis demonstrates that misinformation can be effectively quantified through relation-aware multiclass modeling, supported by curated datasets, scalable annotation strategies, and efficient deep learning architectures.
This thesis makes four primary contributions. First, it provides a comprehensive methodological review of relation extraction, identifying key trends, challenges, and research gaps. Second, it introduces manually curated COVID-19 multiclass relation extraction datasets from Reddit and PubMed, enabling systematic evaluation of transformer-based models, with BioRedditBERT and BioBERT achieving the highest F-scores on the Reddit and PubMed datasets, respectively. Third, the thesis presents COVID-MisinfoRel, a novel manually annotated misinformation dataset that frames misinformation as a multiclass relation extraction problem and demonstrates that transformer models outperform conventional machine learning approaches, such as Random Forest and Support Vector Machine, in distinguishing misinformation categories. Finally, this thesis extends misinformation quantification across multiple domains, including crime, entertainment, finance, and politics, by integrating semi-supervised learning and topic modeling to reduce annotation effort and refine misinformation categories. Furthermore, it presents the first application of the Mamba state-space architecture to quantify misinformation for multiclass relation extraction and also proposes a novel Hybrid Transformer–Mamba model that combines contextual representation learning with efficient sequential modeling to improve scalability and cross-domain generalization.
Overall, this thesis demonstrates that misinformation can be effectively quantified through relation-aware multiclass modeling, supported by curated datasets, scalable annotation strategies, and efficient deep learning architectures.
Kokoelmat
- Väitöskirjat [5374]