PhD in Artificial Intelligence at KInIT: Focused on excellence

The recommender systems are an integral part of almost every modern Web application. Personalized, or at least adaptive, services have become a standard that is expected by users in almost every domain (e.g., news, chatbots, social media, or search).

Obviously, personalization has a great impact on the everyday lives of hundreds of millions of users across many domains and applications. This results in a major challenge – to propose methods that are not only accurate but also trustworthy and fair. Such a goal offers plenty of research opportunities in many directions:

• Novel machine learning approaches for adaptive and recommender systems
• Trustworthy recommendation methods for multi-objective and multi-stakeholder environments
• Explaining recommendations
• Fairness and justice in recommendations
• Biases in the recommendations

There are several application domains where these research problems can be addressed, e.g., search, e-commerce, social networks, news, and many others.

Relevant publications:

• V. Bogina, T. Kuflik, D. Jannach, M. Bielikova, M. Kompan, C. Trattner. Considering temporal aspects in recommender systems: a survey. User Modeling and User-Adapted Interaction, 1-39, 2022. https://doi.org/10.1007/s11257-022-09335-w 
• I. Srba, R. Moro, M. Tomlein, B. Pecher, J. Simko, E. Stefancova, M. Kompan, A. Hrckova, J. Podrouzek, A. Gavornik, and M. Bielikova. Auditing YouTube’s Recommendation Algorithm for Misinformation Filter Bubbles. ACM Trans. Recomm. Syst. 1, 1, Article 6, March 2023. https://doi.org/10.1145/3568392 

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units from abroad. A combined (external) form of study and full employment at KInIT is expected. 

Supervising team

Michal Kompan

Lead researcher, KInIT

Peter Brusilovsky

Professor, University of Pittsburgh, USA

Branislav Kveton

Principal Scientist, Adobe Research, USA

Peter Dolog

Associate Professor, Aalborg University, Denmark

Large language models (LLMs) are increasingly being used for a wide range of downstream tasks where they often show a good performance in zero/few-shot settings compared to specialized fine-tuned models, especially for tasks in which the LLMs can tap into the vast knowledge learned by them during the pre-training. However, they lag behind the specialized fine-tuned models in tasks requiring more specific domain knowledge and adaptation. Additionally, they often suffer from problems such as hallucinations, i.e., outputting coherent, but factually false or nonsensical answers; or generating text laden with biases propagated from pre-training data. Various approaches have recently been proposed to address these issues, such as improved prompting strategies including in-context learning, retrieval-augmented generation or adapting the LLMs through efficient fine-tuning. 

Each of these approaches (or a combination thereof) presents opportunities for new discoveries. Orthogonal to this, there are multiple important factors of models like their level of alignment with human values, their robustness, explainability or interpretability and advances in this regard are welcome as well (generally in AI and particularly in the mentioned approaches).

There are many downstream tasks, where research of the LLM adaptation methods can be applied. These include (but are not limited to) false information (disinformation) detection, credibility signals detection, auditing of social media algorithms and their tendencies for disinformation spreading, and support of manual/automated fact-checking.

Relevant publications:

• Macko, D., Moro, R., Uchendu, A., Lucas, J.S., Yamashita, M., Pikuliak, M., Srba, I., Le, T., Lee, D., Simko, J. and Bielikova, M., 2023. MULTITuDE: Large-Scale Multilingual Machine-Generated Text Detection Benchmark. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9960–9987, Singapore. Association for Computational Linguistics. https://aclanthology.org/2023.emnlp-main.616/ 
• Vykopal, I., Pikuliak, M., Srba, I., Moro, R., Macko, D., and Bielikova, M., 2023. 2024. Disinformation Capabilities of Large Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14830–14847, Bangkok, Thailand. Association for Computational Linguistics. https://aclanthology.org/2024.acl-long.793/ 

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in collaboration with industrial partners or researchers from highly respected research units involved in international projects. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Mária Bieliková

Lead researcher, KInIT

Róbert Móro

Senior researcher, KInIT

The models created in machine learning can only be as good as the data on which they are trained. Researchers and practitioners thus strive to provide their training processes with the best data possible. It is not uncommon to spend much human effort in achieving upfront good general data quality (e.g. through annotation). Yet sometimes, upfront dataset preparation cannot be done properly, sufficiently or at all. 

In such cases the solutions, colloquially denoted as human-in-the-loop solutions, employ the human effort in improving the machine-learned models through actions taken during the training process and/or during the deployment of the models (e.g. user feedback on automated translations). They are particularly useful for surgical improvements of training data through the identification and resolution of border cases. 

Human-in-the-loop approaches draw from a wide palette of techniques, including active and interactive learning, human computation, and crowdsourcing (also with motivation schemes of gamification and serious games). With the recent emergence of large language models (LLM), the original human-in-the-loop techniques can be further boosted to create extensive synthetic training sets with comparatively small human effort. 

The domains of application of human-in-the-loop are predominantly those with a lot of heterogeneity and volatility of data. Such domains include online false information detection, online information spreading (including spreading of narratives or memes), auditing of social media algorithms and their tendencies for disinformation spreading, support of manual/automated fact-checking and more.

Relevant publications:

• Cegin, J., Simko, J. and Brusilovsky, P., 2023. ChatGPT to Replace Crowdsourcing of Paraphrases for Intent Classification: Higher Diversity and Comparable Model Robustness. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing https://arxiv.org/pdf/2305.12947.pdf 
• Šimko, J. and Bieliková, M. Semantic Acquisition Games: Harnessing Manpower for Creating Semantics. 1st Edition. Springer Int. Publ. Switzerland. 150 p. https://link.springer.com/book/10.1007/978-3-319-06115-3 

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Jakub Šimko

Lead researcher, KInIT

Peter Brusilovsky

Professor, University of Pittsburgh, USA

Peter Dolog

Associate Professor, Aalborg University, Denmark

The advent of large language models (LLMs) is raising research questions about how to measure the quality and properties of their outputs. Such measures are needed for benchmarking, model improvements or prompt engineering. Some evaluation techniques pertain to specific domains and scenarios of use (e.g., how accurate are the answers to factual questions in such and such domain? How well can we use the generated answers to train a model for a specific task?), while others are more general (e.g., what is the diversity of paraphrases generated by an LLM? how easy to detect it is that the content is generated?).

Through replication studies, benchmarking experiments, metric design, prompt engineering and other approaches, the candidate will advance the methods and experimental methodologies of LLM output quality measurement. Of particular interest are two general scenarios:

1. Dataset generation and/or augmentation, where LLMs are prompted with (comparatively small) sets of seeds to create much larger datasets. Such an approach can be very useful when dealing with a domain/task with limited availability of original (labelled) training data (such as disinformation detection).
2. Detection of generated content, where stylometric-based, deep learning-based, statistics-based, or hybrid methods are used to estimate whether a piece of content was generated or modified by a machine. The detection ability is crucial for many real-world scenarios (e.g., detection of disinformation or frauds), but feeds back also to research methodologies (e.g., detecting the presence of generated content in published datasets or in crowdsourced data).

The candidate will select (but will not be limited to) one of the two general scenarios, identify, and refine specific research questions and experimentally answer them.

Relevant publications:

• Cegin, J., Simko, J. and Brusilovsky, P., 2023. ChatGPT to Replace Crowdsourcing of Paraphrases for Intent Classification: Higher Diversity and Comparable Model Robustness. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing https://arxiv.org/pdf/2305.12947.pdf 
• Macko, D., Moro, R., Uchendu, A., Lucas, J.S., Yamashita, M., Pikuliak, M., Srba, I., Le, T., Lee, D., Simko, J. and Bielikova, M., 2023. MULTITuDE: Large-Scale Multilingual Machine-Generated Text Detection Benchmark. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing https://arxiv.org/pdf/2310.13606.pdf 

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Jakub Šimko

Lead researcher, KInIT

Dominik Macko

Senior researcher, KInIT

For decades, the behaviour of systems in the physical world has been modelled by numerical models based on the vast scientific knowledge about the underlying natural laws. However, the increasing capabilities of machine learning algorithms are starting to disrupt this landscape. With large enough datasets, they can learn the recurring patterns in the data.

However, a pure machine learning model usually has poor interpretability, needs a lot of data to train which can be hard to come by in many scientific domains, and might not be able to generalize properly. These concerns are being addressed by the field of scientific machine learning (SciML) – an emerging discipline within the data science community. It introduces scientific domain knowledge into the learning process. SciML aims to develop new methods for scalable, robust, interpretable and reliable learning.

Physics-informed neural networks are a part of SciML – we include the physical constraints in the model through appropriate loss functions or tailored interventions into the model architecture. Through physics-informed machine learning, we can create neural network models that are physically consistent, data efficient, and trustworthy.

The goal of the research is to explore how to incorporate scientific knowledge into the machine learning models, thus creating hybrid models based on SciML principles that include both data-driven and domain-aware components. The research could also be directed towards a combination of SciML and transfer learning (that reuses a pre-trained model on a new problem). Such a combination aims to take advantage of both approaches.

SciML can be applied in many domains – we focus mainly on power engineering, e.g. supporting the adoption of renewables and on Earth science with emphasis on positive environmental impact and improving climate resilience, but any other domain could be selected. 

Relevant publications: 

• Kloska, M., Rozinajova, V., Grmanova, G. Expert Enhanced Dynamic Time Warping Based Anomaly Detection. Expert Systems with Applications (2023) https://arxiv.org/pdf/2310.02280.pdf
• Pavlik, P., Rozinajova, V., Bou Ezzeddine, A. Radar-Based Volumetric Precipitation Nowcasting: A 3D Convolutional Neural Network with UNet Architecture. Workshop on Complex Data Challenges in Earth Observation 2022 at CAI-ECAI (2022) https://ceur-ws.org/Vol-3207/paper10.pdf

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units from abroad. A combined (external) form of study and full employment at KInIT is expected. 

Supervising team

Viera Rozinajová

Lead researcher, KInIT

Anna Bou Ezzeddine

Senior researcher, KInIT

Gabriela Grmanová

Senior researcher, KInIT

The recent development of large language models (LLMs) shows the potential of deep learning and artificial neural networks for many natural language processing (NLP) tasks. Advances in automation have a significant impact on a plethora of innovative applications affecting everyday life. 

Although large-scale language models have been successfully used to solve a large number of tasks, several research challenges remain. These may be related to individual natural language processing tasks, application domains, or the languages themselves. In addition, new challenges stemming from the nature of large language models and the so-called black-box nature of neural network-based models. 

Further research and exploration of related phenomena is needed, with special attention to the problem of trustworthiness in NLP or new learning paradigms addressing the problem of low availability of resources needed for learning (low-resource NLP). 

Interesting research challenges that can be addressed within the topic include: 

• Large language models and their properties (e.g., hallucination understanding)
• Trustworthy NLP (e.g., bias mitigation, explainability of models)
• Adapting large language models to a specific context and task (e.g. via PEFT, RAG)
• Advanced learning techniques (e.g., transfer learning, multilingual learning)
• Domain-specific information extraction and text classification (e.g., novel methods for sentiment analysis, improving conversation quality in chatbots)

Relevant publications:

• Pikuliak, M., Hrčková, A., Oreško, Š., Šimko, M. Women Are Beautiful, Men Are Leaders: Gender Stereotypes in Machine Translation and Language Modeling. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 3060–3083, ACL, 2024, https://doi.org/10.18653/v1/2024.findings-emnlp.173

• Pikuliak, M., et al. SlovakBERT: Slovak Masked Language Model. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 7156–7168, ACL, 2022, http://dx.doi.org/10.18653/v1/2022.findings-emnlp.530

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units from abroad. A combined (external) form of study and full employment at KInIT is expected. 

Supervising team

Marián Šimko

Lead researcher, KInIT

Simon Osterman

Senior Researcher & Lead, DFKI

Simeon (Akis) Papadopoulos

Researcher, CERTH

Anders Søgaard

Professor, University of Copenhagen

Martin Hurban

Data Science Team Lead, ČSOB

Large language models (LLM) are a very powerful tool that can be used in a wide range of different applications and they are currently the main driving force of progress in the field of artificial intelligence (AI) for several reasons – e.g. because they help AI systems incorporate a wide range of general knowledge about the world, they can follow natural-language instructions and perform many tasks in few-shot mode, i.e. based on a very small number of examples, thanks to their in-context learning capabilities. They are also able to integrate other modalities (e.g. image and audio).

Despite this unprecedented progress, LLMs also suffer from several significant limitations that currently prevent their wider and safer use in many domains. These restrictions include e.g. the tendency to generate answers that have no support in the training corpus or in the input context (hallucinations), limited ability to perform multi-step reasoning and planning (and apply critical reasoning during training), but also difficulties associated with the integration of other data modalities such as a limited ability to recognize fine-grained visual concepts, etc. LLMs are also much less sample efficient than humans when acquiring new knowledge and skills, which is a significant challenge in some cases – especially for low-resource languages.

This research aims to examine such limitations and – after focusing on one or two of them – propose strategies to mitigate them. Such strategies may include e.g.:

• Developing the ability to perform reasoning e.g. by building upon the bootstrapping reasoning paradigm, adjusting the training paradigm, training on less traditional tasks (e.g. from the reinforcement learning domain), etc.;
• New, more effective self-correction mechanisms and self-evaluation pipelines;
• Improvement of multimodal properties of models, e.g. the ability to recognize fine visual concepts;
• Reducing the rate of hallucinations e.g. by designing new training and fine-tuning techniques, new kinds of LLM pipelines, etc.;
• Mechanisms for reasoning during the training process, supporting the ability to better contextualize the content (e.g. understanding that the text is meant ironically, that it is of lower quality, contains false information, etc.);
• An active training paradigm where models reason and distil during training to acquire new knowledge and skills with improved sample-efficiency;
• …

The application domain might be e.g. support for fact-checking and disinformation combatting, where many of these shortcomings are absolutely critical – but there, of course, is a range of other options.

Relevant publications:

• Srba, I., Pecher, B., Tomlein, M., Moro, R., Stefancova, E., Simko, J. and Bielikova, M., 2022, July. Monant medical misinformation dataset: Mapping articles to fact-checked claims. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (pp. 2949-2959). https://dl.acm.org/doi/10.1145/3477495.3531726
• Pikuliak, M., Srba, I., Moro, R., Hromadka, T., Smolen, T., Melisek, M., Vykopal, I., Simko, J., Podrouzek, J. and Bielikova, M., 2023. Multilingual Previously Fact-Checked Claim Retrieval. https://arxiv.org/abs/2305.07991

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with researchers from highly respected research units. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Michal Gregor

Lead researcher, KInIT

Jana Kosecka

Professor, George Mason University, USA