Scholarly Works
Research Output
Recent Publications
Model-Contingent Polarity Bias in Large Language Model Annotation: Implications for Semantic Multimedia Personalization
Publishing year: 2026
Journal article
Computers, 2026, Volume 15, Issue 5
Repository
Large Language Models (LLMs) are increasingly deployed as automated annotators in semantic multimedia systems, yet their reliability varies significantly across architectures. This study extends prior cross-model evaluations by benchmarking ChatGPT-5, Qwen-3, and Gemini-3-flash against human expert annotations using the HRAST hotel review dataset. We adopt a bias-by-design framework to analyze systematic divergences in sentiment, topic, and aspect labeling across real and synthetic data, while investigating the moderating effects of annotation mode. Findings reveal model-contingent polarity bias: ChatGPT-5 exhibits a pronounced neutrality bias, while Qwen-3 and Gemini-3-flash align more closely with human polarization. Agreement is substantial for concrete topics but diverges on abstract evaluative dimensions. Synthetic data consistently inflates reliability metrics while masking ambiguity. These findings highlight that annotation bias is structurally embedded in model design choices and operational conditions. Cross-architectural triangulation and mode-aware deployment strategies are recommended for robust semantic multimedia system development.
The contemporary aspect of the crime of enforced disappearances
Publishing year: 2025
Conference paper
15th Annual Conference of the European Political Science Association …, 2025, 2025
Google Scholar
Tariff-risk disclosure in 10-Ks and stock market responses to the Liberation Day shock
Publishing year: 2025
Journal article
24th Annual Conference of the Hellenic Finance and Accounting Association (H.F.A.A.), 2025
Repository
This study investigates how firms’ recognition of tariff-related risk in (pre-shock) 10-K disclosures affects stock-market reactions to the April 2, 2025, “Liberation Day” tariff shock. We use a machine-learning approach to identify semantically meaningful tariff risk recognition across the entire 10-K filing, with separate analyses for Risk Factors (“Item 1A”) and the remaining sections of the filing, excluding Item 1A (“Other”). Our results show that firms that explicitly acknowledge tariff risk experience significantly lower abnormal returns around the announcement. Interestingly, this effect is not driven by disclosures in the Item 1A section, but instead arises from tariff-risk recognition in the Other sections of the 10-K and is concentrated among firms facing high litigation risk. The findings highlight the informational role of narrative risk recognition and the importance of where such disclosures appear within corporate filings.
LLMs are not for everything: A Dataset and Comparative Study on Argument Strength Classification
Publishing year: 2025
Conference paper
Proceedings 28th Pan Hellenic Conference on Progress in Computing and Informatics with International Participation PCI 2024, 2025
Scopus
Pages 437-443
A crucial task in computational argumentation is the classification of arguments into categories according to their strength. This work explores the challenge related to this task. It introduces a novel dataset of 615 curated arguments extracted from the CreateDebate platform, chosen for their diversity and relevance to a range of controversial topics. These arguments were annotated for strength into three categories, "weak,""moderate,"and "strong"; Following the annotation process, a rigorous multi-step process was applied to ensure data reliability. The research compares three AI approaches to predict the strength of the argument, a) fine-tuned classification using BERT, b) fine-tuned regression using RoBERTa, and c) prompt-based classification with GPT-4. The results suggest that the regression model significantly outperforms both the fine-tuned classification model and the prompt-based LLM approaches, offering a more accurate and nuanced understanding of the strength of the argument. Despite the results mentioned above, all approaches still face challenges in capturing subtle nuances, highlighting some key areas for future research; Those can include the integration of more sophisticated discourse-aware features, and the development of enhanced representation techniques for argument strength prediction, providing a foundation for future research for automatic argument classification systems.
Improving issue representation with candidate-level voting advice applications
Publishing year: 2025
Journal article
European Journal of Political Research, 2025, Volume 64, Issue 4
Repository
Pages 2132-2145
Voting advice applications (VAAs) have proliferated in recent years. However, most VAAs only match their users with parties, at least in part because creating a VAA matching voters to individual candidates tends to be more labour-intensive. This could be an important missed opportunity. Candidates may deviate from the party line, but voters are often unaware of the policy platforms of individual candidates and therefore rarely hold them accountable for their issue positions in candidate-based elections. VAAs providing information on issue congruence with individual candidates could help to rectify this. We evaluate the potential of candidate-level VAAs by integrating a randomized experiment into a real-world VAA whereby users were exposed either to candidate-level VAA advice or to more standard party-level VAA advice. Our results suggest that candidate-level VAAs are worth the extra effort: they help voters distinguish candidates from parties and cast votes that are more in line with their policy preferences.
Biased by Design? Evaluating Bias and Behavioral Diversity in LLM Annotation of Real-World and Synthetic Hotel Reviews
Publishing year: 2025
Journal article
AI, 2025, Volume 6, Issue 8
Repository
As large language models (LLMs) gain traction among researchers and practitioners, particularly in digital marketing for tasks such as customer feedback analysis and automated communication, concerns remain about the reliability and consistency of their outputs. This study investigates annotation bias in LLMs by comparing human and AI-generated annotation labels across sentiment, topic, and aspect dimensions in hotel booking reviews. Using the HRAST dataset, which includes 23,114 real user-generated review sentences and a synthetically generated corpus of 2000 LLM-authored sentences, we evaluate inter-annotator agreement between a human expert and three LLMs (ChatGPT-3.5, ChatGPT-4, and ChatGPT-4-mini) as a proxy for assessing annotation bias. Our findings show high agreement among LLMs, especially on synthetic data, but only moderate to fair alignment with human annotations, particularly in sentiment and aspect-based sentiment analysis. LLMs display a pronounced neutrality bias, often defaulting to neutral sentiment in ambiguous cases. Moreover, annotation behavior varies notably with task design, as manual, one-to-one prompting produces higher agreement with human labels than automated batch processing. The study identifies three distinct AI biases—repetition bias, behavioral bias, and neutrality bias—that shape annotation outcomes. These findings highlight how dataset complexity and annotation mode influence LLM behavior, offering important theoretical, methodological, and...
69
Unique publications
15.9%
Open access (11 publications; 23 without information)
24.6%
With other institutions (17 publications; 27 without information)
H-Index Reports
| Index type | Scopus | Google Scholar | Web of Science |
|---|---|---|---|
| H-index | 11 | 13 | 1 |
| H5-index | 5 | 6 | 1 |
| H10-index | 6 | 7 | 1 |
Publications per year and source
No description yet.
- Faculty
- Faculty of Communication and Media Studies
- Department
- Department of Communication and Internet Studies