The quality of a teaching method used in a classroom can be assessed by observing the facial expressions of students. To automate this, Facial Expression Recognition (FER) can be employed. Based on the recognized emotions of students, teachers can improve their lectures by determining which activities during the lecture evoke which emotions and how these emotions are related to the tasks solved by the students. Previous work mostly addresses the problem in the context of passive teaching, where teachers present while students listen and take notes, and usually in online courses. We take this a step further and develop predictive models that can classify emotions in the context of active teaching, specifically a robotics workshop, which is more challenging. The two best generalizing models (Inception-v3 and ResNet-34) on the test set were combined with the goal of real-time emotion prediction on videos of workshop participants solving eight tasks using an educational robot. As a proof of concept, we applied the models to the video data and analyzed the predicted emotions with regard to activities, tasks, and gender of the participants. Statistical analysis showed that female participants were more likely to show emotions in almost all activity types. In addition, for all activity types, the emotion of happiness was most likely regardless of gender. Finally, the activity type in which the analyzed emotions were the most frequent was programming. These results indicate that students’ facial expressions are related to the activities they are currently engaged in and contain valuable information for teachers about what they can improve in their teaching practice.

David Dukić, Jan Snajder. Findings of the Association for Computational Linguistics ACL 2024. 2024.

Ana Barić, Laura Majer, David Dukić, Marijana Grbeša-zenzerović, Jan Snajder. Proceedings of the 9th Workshop on Slavic Natural Language Processing 2023 (SlavicNLP 2023). 2023.

This article investigates how the COVID-19 pandemic was framed in public, private and non-profit media production. It conceptualises digital news as an indicator of social resilience and the interaction between social and biological/natural systems. We analysed news articles published in 2020/2021 on 21 Croatian websites using natural language processing. We collected 985,850 articles and manually coded samples to train different classifiers. The first classifier was developed to determine which articles relate to COVID-19. The second classifier was used for articles’ topic classification; the third classifier was used to classify articles into resilience classes. A limited discussion of transformative (long-term) resilience, especially in private media, contributed to the most significant content share. The debate focused on keeping the status quo through coping or returning to pre-pandemic conditions through adaptive mechanisms. The news media contributed to how public issues were framed and how science and scientific research were discussed.

David Dukić, Kiril Gashteovski, Goran Glavaš, Jan Snajder. Findings of the Association for Computational Linguistics: EACL 2024. 2024.

News event monitoring systems allow real-time monitoring of a large number of events reported in the news, including the urgent and critical events comprising the so-called hard news. These systems heavily rely on natural language processing (NLP) to perform automatic event extraction at scale. While state-of-the-art event extraction models are readily available, integrating them into a news event monitoring system is not as straightforward as it seems due to practical issues related to model selection, robustness, and scale. To address this gap, we present a study on the practical use of event extraction models for news event monitoring. Our study focuses on the key task of closed-domain main event extraction (CDMEE), which aims to determine the type of the story’s main event and extract its arguments from the text. We evaluate a range of state-of-the-art NLP models for this task, including those based on pre-trained language models. Aiming at a more realistic evaluation than done in the literature, we introduce a new dataset manually labeled with event types and their arguments. Additionally, we assess the scalability of CDMEE models and analyze the trade-off between accuracy and inference speed. Our results give insights into the performance of state-of-the-art NLP models on the CDMEE task and provide recommendations for developing effective, robust, and scalable news event monitoring systems.