👉The project’s development and activities focus on four key thematic areas: artificial intelligence for the green transition, for health and biotechnology, for the digital society, and for science. The goal is to actively promote collaboration between industry, academia, and research institutions, create opportunities for joint projects, and enable the transfer of knowledge and technologies into practice.

Video

Recording is already available.

Methods for semi-automated hypothesis generation from scientific literature: an open science approach

The rapid growth of scientific publications makes it difficult to manually review and keep up to date with new research findings. Literature-based discovery (LBD) is a field of artificial intelligence at the intersection of natural language processing and machine learning, which enables semi-automated hypothesis generation by discovering new associations between previously unconnected scientific sources. In the lecture, we will present a selection of approaches and methods from our recently published monograph Bisociative Literature-Based Discovery: Methods with Tutorials in Python (Springer, 2025). We will present also a collection of Python notebooks that facilitate the reproducibility of procedures for data acquisition, text processing, hypothesis generation and their evaluation, in alignment with the principles of open science.

About the lecturers: Nada Lavrač is a scientific councilor at the Jožef Stefan Institute and a full professor at the Jožef Stefan International Postgraduate School. Bojan Cestnik is the director of the computer engineering company Temida, a researcher at the Jožef Stefan Institute and a full professor at the School of Engineering and Management of the University of Nova Gorica. Andrej Kastrin is a researcher at the Institute for Biostatistics and Medical Informatics at the Faculty of Medicine, University of Ljubljana, and a lecturer at this faculty.

The open day was intended to establish a dialogue with companies so that SLAIF could co-design its services according to their needs. The event confirms that the Slovenian economy sees artificial intelligence as a key opportunity for competitiveness, innovation and digital transformation.
SLAIF will continue to provide access to top-notch infrastructure, expertise and support, thus promoting closer connections between science and industry.

Prof. Dr. Sašo Džeroski, Head of the Department of Knowledge Technologies, highlighted the long-standing excellence of the Slovenian research area in the field of artificial intelligence. He stressed that Slovenia is particularly strong in the use of AI in science. He mentioned key JSI European projects and the SLAIF project – the Slovenian Artificial Intelligence Factory, which connects researchers with the economy and accelerates knowledge transfer. Prof. Dr. Džeroski emphasized that the scientific community, in addition to stable funding, also needs close cooperation with the government and ministries. Only in this way can development priorities be aligned with the actual needs of society, the economy, education and public services. The Jožef Stefan Institute will continue to actively contribute to Slovenia’s breakthrough in the field of artificial intelligence.

Prof. Dr. Dzeroski co-organized and moderated the Planet & Climate workshop together with Prof. Dr. Markus Reichstein, Director at the Max Planck Institute for Biogeochemistry. The workshop focused on the role of AI and machine learning in advancing planet and climate science, with discussions on weather prediction, climate mitigation, Earth observation, and innovation in environmental technologies.

AIS25 gathered more than 1,200 delegates from over 50 countries, fostering discussions on the transformative role of AI in scientific discovery. The launch of the RAISE initiative further highlighted Europe’s commitment to strengthening collaboration and coordination of AI research, infrastructure, and expertise across Europe.

Congratulations!

Abstract:

Face morphing attacks pose a significant threat to biometric security systems by enabling multiple individuals to authenticate with a single compromised credential i.e., a morphed face image. This thesis investigates the use of multimodal large language models (MLLMs) for morphing attack detection, demonstrating that foundation models trained on large-scale, heterogeneous data possess latent forensic capabilities that can be adapted for specialized security tasks.

We evaluate four open-source models in a zero-shot setting, including Gemma- 3 27B, Qwen2.5-VL 32B, Llama-4 Scout 17B, and Mistral Small 3.1 24B, across diverse datasets covering landmark-based, GAN-based, and diffusion-based morphing attacks. Even without task-specific training, these models achieve measurable detection performance, confirming that multimodal language models inherently encode useful representations. To improve zero-shot detection reliability, we developed a structured forensic prompt, which guides the models through a systematic six-step procedure for detecting visual artifacts created during the blending of facial images. This structured prompting approach enhances both detection accuracy and interpretability of the outputs.

The primary contribution of the thesis lies in parameter-efficient fine-tuning through Low-Rank Adaptation (LoRA). Using only 0.61% of trainable parameters, we fine-tuned Gemma-3 12B. This fine-tuned model substantially outperformed its zero-shot counterpart, reducing the average Equal Error Rate by more than half. It achieved near-perfect detection on landmark-based morphs, competitive results on challenging GAN-based and diffusion-based morphs. Overall, this research establishes multimodal large language models as a viable and promising direction for morphing attack detection, combining generalization and interpretability with competitive performance against state-of-the-art approaches.

About

This monograph introduces the field of bisociative literature-based discovery (LBD) by first explaining the underlying LBD principles and techniques, followed by the presentation of bisociative LBD techniques and applications developed by the authors. LBD is a process of uncovering new knowledge by analyzing and connecting disparate pieces of information from different sources of literature.

Selected techniques include conventional natural language processing (NLP) approaches, as well as outlier-based, concept-based, network-based, and embeddings-based LBD approaches. Reproducibility aspects of bisociative LBD research are also covered, addressing all steps of the bisociative LBD process: data acquisition, text preprocessing, hypothesis discovery, and evaluation.

The monograph is targeted at researchers, students, and domain experts interested in knowledge exploration, information retrieval, text mining, data science or semantic technologies. By covering texts, relations, networks, and ontologies, this work empowers domain experts to transcend their knowledge silos when confronted with varied data formats in their research practice. The monograph’s open science approach with tutorials in Python allows for code reuse and experiment replicability.

Keywords:

• text mining
• document representation
• bisociative knowledge discovery
• outlier detection
• graph mining
• literature-based discovery

Congratulations!

Abstract:

Bug severity classification is a critical and time-consuming aspect of the software bug resolution process, as the severity of a reported bug influences its urgency to fix it. This task is often automated using supervised machine learning methods; however, there is a notable lack of dedicated lexicons for this purpose. In this thesis, we assessed the potential usefulness of developing such a resource, implemented it and compared its performance with classic machine learning approaches and general purpose lexicons used in text classification. We also proposed and tested a novel lexicon development approach based in the domain of software bug reports. In our empirical assessment, we used publicly available datasets of bug reports for Firefox and Eclipse. The results indicate that our lexicon approach achieved a comparable F1-score to classic machine learning models. Our experiments revealed that lexicon-based approaches were effective in most of the experiments, while machine learning methods performed better in the experiment that had a more balanced data distribution. The findings highlighted the dataset-dependent nature of classification performance and also indicated the unexpected usefulness of general sentiment analysis lexicons. These results suggest that tailored lexicon-based approaches are a valuable alternative to machine learning techniques for bug severity classification and could potentially reduce the need for large labeled training sets.

Congratulations!

Abstract:

This thesis presents a computational analysis of socially biased and dehumanising discourse in Slovene news media, using natural language processing techniques. Social biases are not only reflected, but also perpetuated in language, and dehumanising discourse both results from and results in a discriminative perception and treatment of a specific social group. Specifically, we focus on the discourse on migrants in Slovene news media in the migration periods following the wars in Syria (2015-2016) and in Ukraine (2022-2023), and on the
representation of migrants and members of the LGBTQIA+ community in news media consumed by a left-,centre, or right-wing-leaning public.

The main contribution of the thesis is a novel adaptation and application of natural language processing techniques to the detection of social bias and dehumanisation in Slovene. The approaches employed include the training of static word embeddings, vector similarity, sentiment analysis, and masked token prediction.

The results of the empirical studies reveal that Slovene news articles about migrants are generally more negative, intense, and dehumanising during the migration period following the war in Ukraine compared to the period following the war in Syria. For instance, migrants are more closely associated with concepts such as moral disgust and vermin during the period of the Ukrainian war compared to the Syrian war. However, when comparing the articles about Ukrainian migrants and other migrants in the 2022-2023 period, the
ones specifically referring to Ukrainian migrants are more positive, more intense, and less dehumanising than those referring to other migrants. Furthermore, female migrants and female members of the LGBTQIA+ community face higher levels of dehumanisation in
media outlets consumed by a right-wing public compared to those read by a centrist or left-wing audience.

The results emphasise the urgent need for continued research to address the harmful effects of socially-biased and dehumanising language in news media.