This study introduces a benchmark for evaluating information retrieval in the Slovak language, leveraging a question-answering dataset for fine-tuning and assessment of sentence transformers. The dataset, named Retrieval SkQuAD, is integrated into two widely recognized evaluation frameworks: BEIR (Benchmarking Information Retrieval) and MTEB (Massive Text Embedding Benchmark). Retrieval SkQuAD (Slovak Question Answering Dataset) comprises 19,000 manually annotated answers to 1,134 questions, with each answer assigned a relevance score, and includes information on the usefulness of documents in generating responses. Unlike question answering datasets, this resource provides a nuanced assessment of partial relevance across multiple documents. We fine-tuned several sentence transformers and BERT-based models specifically for retrieving documents containing correct answers within the Slovak Wikipedia. Our fine-tuning process incorporates adversarial questions as hard negatives, leading to significant improvements in retrieval accuracy. Experimental results demonstrate that our approach advances state-of-the-art performance for Slovak information retrieval.

This paper presents BIANCA (Business Intelligence and ANalytics through Cognitive Assessment), a novel multi-agent framework designed for collaborative incident investigation in IT Service Management (ITSM) environments. BIANCA leverages Large Language Models (LLMs) orchestrated through specialized agent collaboration to automate complex root cause analysis workflows by integrating heterogeneous operational data as ITSM tickets, CMDB/configuration repositories, enterprise documentation and wiki knowledge that transform them to unified knowledge graph representation of applications, systems, network components and their dependencies. Large Language Models (LLMs) are orchestrated as specialized agents that query these sources, consume the graph-derived context, and synthesize root cause hypotheses and operator-oriented troubleshooting guidance. The framework implements a node-centric payload deduplication algorithm that achieves 60–70

We introduce SkMTEB, the first comprehensive MTEB-style text embedding benchmark for Slovak, a low-resource West Slavic language, comprising 31 datasets across 7 task types—nearly 4 × the depth of existing multilingual benchmark coverage for Slovak. Our evaluation of 31 embedding models reveals that large instruction-tuned multilingual models achieve the strongest performance, while existing Slovak-specific models trained for NLU tasks transfer poorly to embedding tasks. To address the need for efficient, locally-deployable Slovak embeddings, we develop e5-sk-small (45M parameters) and e5-sk-large (365M) by applying vocabulary trimming and fine-tuning to Multilingual E5 models. Despite size reductions of up to 62%, our open-source models achieve competitive performance with proprietary APIs while remaining locally deployable for semantic search and retrieval-augmented generation (RAG). We release the benchmark, models, datasets, and code openly, hoping our approach offers a replicable path for other under-resourced languages.

Identifying a person's age and gender from speech signal characteristics poses a significant challenge in personal identity recognition systems, particularly when security considerations are involved. In signal processing applications such as speaker recognition, biometric identification, human-machine interface (HMI), and telecommunication, the estimation of age and gender from voice is a crucial and demanding problem. In several signal processing domains, deep learning models have demonstrated remarkable effectiveness. In this paper, we propose a modified convolutional neural network to identify the age and gender of the speaker using the characteristics of the MFCC speech. We also included techniques to reduce the dimensionality of the speech feature set. We tested modified one-dimensional convolutional neural networks (1D-CNN) and machine learning models such as support vector classification (SVC), decision trees (DT), and random forests (RF). The modified 1D-CNN based on deep learning, along with dimensionality reduction, random seeding, and cross-validation, is proposed for the recognition of age and gender in speech. We applied different dimensionality reduction techniques, such as principal component analysis (PCA) and independent component analysis (ICA), along with random seeding and various sets of cross-validation. In this study, we used the Children Speech Recorning Dataset, Biometric Visions and Computing (BVC), and the Mozilla Common Voice speech datasets for estimating age and gender from speech. The proposed 1D-CNN model exhibits a promising performance compared to the state-of-the-art (SOTA) approaches. The models were evaluated and compared with evaluation metrics, such as accuracy. The dimensionality reduction techniques, selection of speech features, and seeding show a significant impact on the performance of the suggested model.

In this work, we introduce skLEP, the first comprehensive benchmark specifically designed for evaluating Slovak natural language understanding (NLU) models. We have compiled skLEP to encompass nine diverse tasks that span token-level, sentence-pair, and document-level challenges, thereby offering a thorough assessment of model capabilities. To create this benchmark, we curated new, original datasets tailored for Slovak and meticulously translated established English NLU resources. Within this paper, we also present the first systematic and extensive evaluation of a wide array of Slovak-specific, multilingual, and English pre-trained language models using the skLEP tasks. Finally, we also release the complete benchmark data, an open-source toolkit facilitating both fine-tuning and evaluation of models, and a public leaderboard at https://github.com/slovak-nlp/sklep in the hopes of fostering reproducibility and drive future research in Slovak NLU.

This study explores the enhancement of Slovak Natural Language Processing (NLP) by fine-tuning Sentence BERT models using the translated MS MARCO V2 dataset, facilitated by the MADLAD-400 language model. We translated the MS MARCO V2 dataset into Slovak and utilized this translation to train Sentence BERT models, aiming to improve the quality of vector representations for Slovak sentences. Employing a triplet network approach, we derived meaningful sentence vector representations, enabling effective comparison through cosine similarity. The research yields a newly translated database and evaluates the performance of fine-tuned Slovak vector representation models, thereby advancing Slovak search capabilities.

The limited availability of language resources for Slovak presents a significant challenge for the development and evaluation of language models. In this paper, we introduce a multiple-choice question-answering dataset specifically designed for the financial domain in Slovak. The dataset contains 1,334 questions, each with one correct answer and four incorrect ones. It is systematically organized by topic and difficulty level to facilitate structured evaluation. Using this dataset, we assess the performance of several Slovak generative language models and compare their results against a general question-answering dataset to analyze domain-specific model capabilities. The best-performing model is a monolingual Slovak model. Furthermore, the observed performance differences between financial-domain and general question-answering tasks suggest that domain-specific language modeling requires further research.

In today's digital landscape, the widespread use of large language models has made it increasingly difficult to differentiate between human-written and AI-generated text. This challenge arises because state-of-the-art LLMs like Gemini, GPT series including ChatGPT GPT-4, and LLaMa produce highly sophisticated, human-like text. This indistinguishability poses various problems across various sectors, including cybersecurity threats, the generation of propaganda, the dissemination of biased or false information on social media, and facilitating social engineering attacks. In the educational domain, these models contribute to academic dishonesty, while in complex, multi-team environments, they add new layers of intricacy in managing human-AI interactions. All source code is available in my GitHub repository (https://github.com/mohanreddy91/AI-Generated-Text-Detection.).

The rise of social media has led to an increase in toxic language, hate speech, and offensive content. While extensive research exists for widely spoken languages like English, Slovak remains underrepresented due to the lack of high-quality datasets. This gap limits the development of effective models for toxicity detection and sentiment analysis in Slovak. To address this, we introduce three new annotated Slovak datasets focused on toxic language, offensive language, hate speech detection, and sentiment analysis. These native datasets provide a more reliable foundation for automated moderation compared to machine-translated alternatives. Our research also highlights the real-world impact of online toxicity, including social polarization and psychological distress, emphasizing the need for proactive detection systems on social media platforms. This paper reviews existing Slovak datasets, presents our newly developed resources, and provides a comparative analysis. Finally, we outline key contributions and suggest future directions for improving toxic language detection in Slovak.

During semiconductor manufacturing, wafer defect patterns emerge in an uncontrolled environment, making immediate recognition challenging. To enhance the classification accuracy in pattern recognition, deep learning (DL) techniques are employed to address the issue of defective pattern identification. The inherent differences among various wafer defect types lead to imbalanced testing datasets, thereby reducing the efficiency of classification models in delivering accurate predictions. This study proposes a convolutional neural network (CNN) combined with a data transformation method at the feature extraction stage, incorporating Density-based features and Radon-based features to enhance the representation of key patterns. Subsequently, multiple classifiers are integrated using voting, bagging, and AdaBoost strategies to improve classification performance and model generalization. The primary objective is to mitigate data imbalance and selection bias by employing random sampling and assigning distinct weights to individual classifiers. The contributions of this study are threefold. First, we introduce a novel feature extraction strategy that leverages Features extracted by the CNN model, Density-based features and Radon-based features, effectively capturing structural and spatial characteristics in wafer defect patterns. Second, we propose an ensemble learning framework integrating multiple classifiers with optimized weighting mechanisms to enhance classification robustness. Third, we provide empirical evidence demonstrating that our weighted soft voting approach achieves superior performance, attaining a classification accuracy of 95.09% and an F1 score of 0.95. These results validate the effectiveness of our method in improving the reliability of wafer defect classification, which is crucial for advancing automated defect inspection in semiconductor manufacturing.

Data privacy is crucial today, especially with regulations such as GDPR. Data anonymization is key, but common methods often reduce data value. This paper explores the acronymization process, which replaces sensitive data with abbreviations, as a way to balance protection and data usability. We propose a new approach: using large language models (LLMs) to create training data for a T5 model, which we then fine-tune for acronymizing sensitive data. The results show that this method, especially when using LLMs like Gemma-9B-IT to augment the data, achieves promising results and outperforms existing Named Entity Recognition (NER) models in the specific task of acronymization. This offers a more efficient and scalable solution for anonymizing text data, contributing to both privacy protection and preserving the utility of data for analysis and research.

Dementia is a neurodegenerative disorder that affects the brain and causes a decrease in cognitive abilities that interfere with memory, thinking, and the ability to perform daily activities. Detecting dementia automatically is a difficult task and involves a detailed analysis of speech and language characteristics and an examination of the participant’s mental state. The picture description task is one of the screening tools for dementia detection from transcribed speech, where participants are instructed to describe what they see in a picture. In this paper, we evaluate the performance of available monolingual and multilingual pre-trained BERT models fine-tuned on transcribed Slovak speech from the picture description task with the aim of automatic dementia detection. The results of the F1-score range from 69% to 93% in the binary classification task and are comparable with recently published studies in this area.

This paper presents a comparative analysis of clas-sifiers for detecting grammatical errors in Slovak sentences. We experiment with state-of-the-art transformer models, in-cluding the BERT multilingual base model, SlovakBERT, and XLM-RoBERTa, trained on datasets containing synthetically introduced grammatical errors. The datasets were created by corrupting words in various ways, simulating different types of grammatical errors such as incorrect letters, punctuation, and case sensitivity. These datasets range in size from 50,000 to 1,000,000 sentences, with varying ratios of correct-to-incorrect sentences. The models were fine-tuned on these datasets and evaluated using key metrics such as accuracy, precision, recall, and F1 score. Our results highlight the impact of dataset size, error type, and sentence distribution on the models' ability to detect grammatical errors. The findings provide valuable insights into the development of grammar-checking systems for the Slovak language.

Grammatical error correction is a key challenge in natural language processing research, driving advancements in language processing. It ensures accuracy and polish in written content for authors and creators across domains. In this paper, we evaluate datasets focused on grammatical error correction in the Slovak language using the Text-to-Text Transfer Transformer model. We have outlined significant studies within the realm of natural language processing, as well as the utilization of the T5 model in related research. In our experimental approach, we employed two datasets sourced from various online platforms. Our endeavors yielded remarkable outcomes both within this domain and specifically for the Slovak language.

This paper presents a comprehensive evaluation of three language models: RoBERTa, SlovakBERT, and BERT-Multilingual, using datasets of varying sizes (ranging from 50,000 to 1,000,000 sentences) to assess their performance in natural language processing tasks in the Slovak language. The evaluation focuses on four critical metrics: accuracy, precision, recall, and F1 score across different dataset distributions (70:30, 50:50, and 30:70). The results demonstrate SlovakBERT's superior performance across nearly all metrics and distributions, particularly excelling in precision and F1 scores, which suggests its robustness and reliability in handling varied data distributions effectively. RoBERTa, while showing variability, performs well in precision under conditions of data imbalance but faces challenges in maintaining consistent accuracy. BERT-Multilingual, though slightly trailing behind, still delivers solid performance, especially in precision, making it a dependable model across varied conditions. The findings underscore the effectiveness of SlovakBERT in scenarios requiring high precision and reliability, and provide insights into the strengths and limitations of each model under different testing conditions.

This paper provides a review of the GPT-J demonstration with a specific focus on its fine-tuning capabilities for text generation in Slovak. Through reasoning, we explore the model’s fine-tuning in tasks such as text generation. We compare the results between different sizes of datasets and different sizes of pre-trained models. The results clearly show that the size of the dataset has an impact on the quality of the resulting model.

The process of determining and understanding the emotional tone expressed in a text, with a focus on textual data, is referred to as sentiment analysis. This analysis facilitates the identification of whether the overall sentiment is positive, negative, or neutral. Sentiment analysis on social networks seeks valuable insight into public opinions, trends, and user sentiments. The main motivation is to enable informed decisions and an understanding of the dynamics of online discourse by businesses and researchers. Additionally, sentiment analysis plays a vital role in the field of hate speech detection, aiding in the identification and mitigation of harmful content on social networks. In this paper, studies on the sentiment analysis of texts in the Slovak language, as well as in other languages, are introduced. The primary aim of the paper, aside from releasing the “SentiSK” dataset to the public, is to evaluate our dataset by comparing its results with those of other existing datasets in the Slovak language. The “SentiSK” dataset, consisting of 34,006 comments, was created, specified, and annotated for the task of sentiment analysis. The proposed approach involved the utilization of three datasets in the Slovak language, with nine classification methods trained and compared in two defined tasks. For the first task, testing on the “SentiSK” and “Sentigrade” datasets involved three classes (positive, neutral, and negative). In the second task, testing on the “SentiSK”, “Sentigrade”, and “Slovak dataset for SA” datasets involved two classes (positive and negative). Selected models achieved an F1 score ranging from 75.35% to 95.04%.

Deep learning (DL) methods are widely employed in the semiconductor manufacturing process to enhance pattern recognition and classification accuracy, specifically for addressing defect patterns. However, the classification performance of the current models is hindered by the imbalanced distribution of defect data within the test dataset. To tackle this issue, this study presents a feature extraction approach utilizing data transformation, and ensemble learning techniques aiming to enhance the model's classification performance. The primary objective of this study is to mitigate selection and imbalance problems in the dataset through random sampling and assigning distinct weights to individual classifiers. The results demonstrate that the proposed method achieves an impressive accuracy rate of 95.09%, thus substantiating its efficacy in improving the robustness of both the classification model and wafer classification.

Automatic detection of emotional stress in speech has recently drawn increasing attention, mainly in the fields of natural and spoken language processing. Stress can be automatically detected from speech-derived parameters, such as fundamental frequency F0, voice intensity, jitter, shimmer, harmonic-to-noise ratio, mel-frequency cepstral coefficients, etc. In this paper, we focus on the analysis of selected voice characteristics in available datasets of emotional speech and speech under stress in the English and Slovak languages. We also proposed a stress level detection approach, which uses the fusion of the Teager energy operator (TEO) with mel-frequency cepstral coefficients (MFCCs) and classification based on decision trees and ensemble learning method of bagging decision trees. Preliminary results of 64% classification accuracy are comparable to other studies in this field of research.