Six New Articles Published
Christopher Bagdon, Carina Silberer and Roman Klinger
Emotion experience, expression, and perception: Emotion analysis on multimodal social media posts. In The 2026 Conference on Empirical Methods in Natural Language Processing, 2026.
In this study, we found that people only correctly understand the emotions expressed in social media posts if they are also familiar with the background to the situation described. If important information is missing, or if the sequence of events leading up to the post is misunderstood, the emotion expressed is often misinterpreted as well. The context of a post is therefore crucial to how we perceive the emotions of others online.
-
Jiahui Li, Yarik Menchaca Resendiz, Sean Papay and Roman Klinger
iPOE: Interpretable prompt optimization via explanations. In The 2026 Conference on Empirical Methods in Natural Language Processing, 2026.
A new method has been developed that enables prompts to be improved in a targeted and traceable manner. Clear guidelines help the language model to make better decisions and make the optimisation process transparent. This enables even people with no prior knowledge to create good prompts, whilst at the same time achieving significantly better results than with previous methods.
-
Albina Sarymsakova and Roman Klinger
Are more emotion theories better? Joint prompting improves emotion categorization robustness. In The 11th Workshop on Natural User-generated Text, 2026.
The new study examines how reliably large language models can recognise emotions in texts. The results show that when the models are instructed to combine several psychological theories of emotion, they produce more reliable predictions even when dealing with distorted or flawed texts. This approach can help to develop more robust systems for automatic emotion recognition.
-
Sabine Weber, Angelina Wang, Ankush Gupta, Arjun Subramonian, Dennis Ulmer, Eshaan Tanwar, Geetanjali Aich, Hannah Devinney, Jacob Hobbs, Jennifer Mickel, Joshua Tint, Mae Sosto, Ray Groshan, Simone Astarita, Vagrant Gautam, Verena Blaschke, William Agnew, Wilson Y Lee, Yanan Long
Queer NLP: A Critical Survey on Literature Gaps, Biases and Trends. In: Transactions of the Association for Computational Linguistics, 2026.
This new review study examines research on LGBTQIA+ communities and language technologies. Although the number of papers on ‘Queer NLP’ is growing, so far these have tended to highlight existing prejudices and problems rather than developing concrete solutions. There is a particular need for research into the involvement of affected groups, the consideration of different forms of discrimination, collaboration between different disciplines, and the study of languages other than English.
-
Itisha Yadav, Roman Klinger and Diana Peters
Where, how, and what to ground: Evaluating ontology-guided llm-based knowledge graph construction. In 5th International Workshop on Natural Language Processing for Knowledge Graph Construction (NLP4KGC 2026), co-located with ISWC 2026, 2026.
This study investigated how large language models can transfer information from documents into knowledge graphs using ontologies – that is, structured conceptual frameworks. Experiments with safety data sheets show that the results are better when the ontology is incorporated during the information extraction process, rather than the data being mapped retrospectively. Ontologies that are specifically tailored to the relevant subject area produce particularly precise and detailed representations. These findings will help to ensure that even large and complex industrial ontologies can be used more reliably in AI-supported systems in future.
-
Aswathy Velutharambath, Kai Sassenberg and Roman Klinger
What if deception cannot be detected? A cross-linguistic study on the limits of deception detection from text.Computational Linguistic, pages 1--72, 07 2026.
The study investigates whether deception can be detected solely on the basis of written texts. To this end, it introduces an approach that defines deception as a contradiction between a statement and the writer’s actual belief, and creates three datasets named DeFaBel. The linguistic features examined show no statistically significant correlation with deception; even automated detection models perform barely better than chance. Previous successes may therefore be due to specific characteristics of the respective datasets and cannot be reliably generalised.