Margit Antal

Embedding models have become a fundamental component of modern natural language processing, yet their performance in morphologically rich, low-resource languages such as Hungarian remains underexplored. In this paper, we present a systematic evaluation of state-of-the-art embedding models for Hungarian question-answer retrieval. We construct two complementary evaluation datasets: (i) a domain-specific corpus collected from company documentation, preprocessed into topical chunks with human-verified question-answer pairs, and (ii) the publicly available HuRTE benchmark. Using FAISS as the vector database, we compare eight multilingual and crosslingual embedding models. Performance is measured using Mean Reciprocal Rank (MRR) and Recall@k. Results show substantial variation across models and datasets, with notable differences between domain-specific and general-purpose retrieval tasks. We complement the evaluation with error analysis, highlighting challenges posed by Hungarian morphology and compounding, and discuss trade-offs in efficiency. Our findings provide the first comparative study of embedding-based retrieval in Hungarian, offering practical guidance for downstream applications and setting a foundation for future research in Hungarian representation learning. The dataset and the corresponding evaluation code are publicly accessible at https://github.com/margitantal68/ hu-embeddings-hurte.

Retrieval-Augmented Generation (RAG) systems have gained significant attention for their ability to enhance language models by incorporating external knowledge sources. However, evaluating the effectiveness of these systems remains a challenge, as both the retrieval and generation components must be assessed independently and in combination. This study aims to provide a comprehensive evaluation of open-source large language models (LLMs) within a RAG system, using a dataset derived from diploma theses abstracts. To systematically assess performance, we generated a benchmark dataset consisting of 122 questions with reference answers, categorized into reasoning, fact-based, and summary-type questions. Using Elasticsearch as the retriever and several open-source LLMs as generators, we evaluated the system using the Ragas (Retrieval Augmented Generation Assessment) framework, focusing on retrieval effectiveness and answer quality. Our findings highlight the strengths and weaknesses of different models and retrieval strategies, offering insights into optimizing RAG implementations for academic and structured knowledge tasks. Our dataset and the corresponding evaluation code are publicly accessible at https://github.com/margitantal68/rag_paper .

This study investigates the personality traits of large language models (LLMs) using the Eysenck Personality Questionnaire (EPQ) and compares their responses to those of human participants. By administering the EPQ three times to both groups, we aim to analyze the stability and consistency of their responses over time. Our research focuses on open-source LLMs. The results indicate that LLMs exhibit significantly higher extraversion scores compared to humans, while differences in lie, neuroticism, and rigidity traits are not statistically significant. Additionally, LLMs show greater variability in their responses across repeated administrations, suggesting less consistency compared to human participants. We also examine the effects of prompt design and the temperature hyperparameter on LLM personality traits. Both temperature settings and prompt instructions significantly shape the personality-like patterns in LLM outputs. These findings contribute to the understanding of how LLMs can mimic human personality traits and the implications for their use in personalized communication and user engagement.

In this paper, we report on PIN-code typing behaviour on touchscreen devices of 112 subjects. Detailed statistical analysis revealed that the major di erence between subjects is in inter-key latency. Key-press duration variations are insignificant compared to inter-key latency variations. Subjects were grouped into meaningful clusters using clustering. The resulting clusters were of slow, medium, and fast typists. The dataset was split randomly into two equal size subsets. The first subset was used to train different synthetic data generators, while the second subset was used to evaluate an authentication attack using the generated synthetic data. The evaluation showed that slow typists were the hardest to attack. Both the dataset and the software are publicly available at https://github.com/margitantal68/sapipin_paper.

The growing interest in bot detection can be attributed to the fact that fraudulent actions performed by bots cause surprisingly high economical damage. State-of-the-art bots aim at mimicking as many as possible aspects of human behavior, ranging from response times and typing dynamics to human-like phrasing and mouse trajectories. In order to support research on bot detection, in this paper, we propose an approach to generate human-like mouse trajectories, called SapiAgent . To implement SapiAgent, we employ deep autoencoders and a novel training algorithm. We performed experiments on our publicly available SapiMouse dataset which contains human mouse trajectories collected from 120 subjects. The results show that SapiAgent is able to generate more realistic mouse trajectories compared with Bézier curves and conventional autoencoders.

Behavioural biometrics provides an extra layer of security for user authentication mechanisms. Among behavioural biometrics, mouse dynamics provides a non-intrusive layer of security. In this paper we propose a novel convolutional neural network for extracting the features from the time series of users’ mouse movements. The effect of two preprocessing methods on the performance of the proposed architecture were evaluated. Different training types of the model, namely transfer learning and training from scratch, were investigated. Results for both authentication and identification systems are reported. The Balabit public data set was used for performance evaluation, however for transfer learning we used the DFL data set. Comprehensive experimental evaluations suggest that our model performed better than other deep learning models. In addition, transfer learning contributed to the better performance of both identification and authentication systems.