23/2026

Online archive of Computer Science and Mathematical Modelling

23/2026

  1. Szymon Florek, Piotr Górny - Ablation study of correction methods in Polish language Retrieval-Augmented Generation chatbots
    Pages: 5 - 18
    Abstract: Retrieval-Augmented Generation (RAG) mitigates hallucinations in large language models by grounding responses in retrieved evidence, yet the interplay of correction techniques across the RAG pipeline remains underexplored – especially for non-English languages. We present a modular RAG system evaluated through a systematic ablation study of 12 pipeline configurations on a 1,000-question Polish-language test set derived from a 1M-article Wikipedia corpus and manually validated. Four correction families are compared: (i) cross-encoder reranking, (ii) query decomposition with multi-hop reranking and diversity enforcement, (iii) native Chain-of-Thought (CoT) reasoning, and (iv) post-generative Chain-of-Verification (CoVe). Results evaluated with RAGAS metrics show that multi-hop reranking yields the largest quality gains (Δ Faithfulness +15.6%, Δ Context Recall +38.4%, Δ Relevancy +20.3%) at acceptable latency (18.4 s), while CoVe in auto-correction mode can paradoxically degrade faithfulness at high computational cost (+41.6 s). Cross-encoder reranking achieves the best cost quality ratio (+16.7% Recall, +1.2 s). We discuss practical configuration tiers for real- -time, interactive, and offline deployments.
    Keywords Retrieval-Augmented Generation, ablation study, multi-hop reranking, Chain-of-Thought, Chain-of-Verification, Polish NLP, chatbot evaluation
    Full article: PDF icon 5_18_sflorek_pgorny_csmm_23_2026.pdf
  2. Piotr Gdula - Real-time sign language to speech translation using CNNs
    Pages: 19 - 27
    Abstract: This article presents the design, implementation, and empirical evaluation of an end-to-end software architecture for the real-time translation of the American Sign Language (ASL) alphabet into text and synthesized speech. The proposed system leverages the OpenCV library for video ingestion and the MediaPipe framework to extract 21 precise three-dimensional topological landmarks of the human hand, effectively isolating the gesture from environmental noise. A custom Convolutional Neural Network (CNN) built in TensorFlow and Keras processes these normalized skeletal representations to classify 26 static ASL characters. The model demonstrates a robust overall classification accuracy, performing exceptionally well on open hand topologies while highlighting occlusion-based challenges in complex gestures. Operating with exceedingly low latency, the system ensures fluid real-time human-computer interaction, providing a scalable and computationally efficient foundation for accessible communication tools without the need for specialized depth-sensing hardware.
    Keywords sign language translation, Convolutional Neural Networks, real-time processing
    Full article: PDF icon 19_27_pgdula_csmm_23_2026.pdf
  3. Krzysztof Maj - Use of generative artificial intelligence models for image generation in information operations (InfoOps)
    Pages: 29 - 38
    Abstract: This paper analyzes the use of generative artificial intelligence (GenAI) models for image generation in the context of information operations (InfoOps). The study adopts a qualitative approach based on literature review and case study analysis of selected examples of synthetic visual content used in disinformation campaigns. The paper examines key components of this phenomenon, including threat actors, technologies, information influence vectors, and distribution channels. An original threat model is proposed to structure the mechanisms through which synthetic images are created, disseminated, and leveraged for influence. The findings indicate that GenAI significantly enhances the scalability, speed, and effectiveness of information operations by enabling the rapid production of highly realistic visual content. The analysis also demonstrates that the impact of synthetic images depends not only on their quality but on their integration within broader communication ecosystems. The study identifies key risks, including manipulation of public opinion, erosion of trust, and increased information disorientation. The results highlight the growing importance of interdisciplinary approaches to counter disinformation, including the development of detection methods and the strengthening of critical media literacy.
    Keywords generative artificial intelligence, image generation, information operations
    Full article: PDF icon 29_38_kmaj_csmm_23_2026.pdf
  4. Wiktoria Sakowska - Automating business document processing: a comparative review of OCR, machine learning and large language models
    Pages: 39 - 45
    Abstract: Automating business document processing requires more than character recognition: fields must be identified, layout relations preserved and the structured output validated. This paper reviews four families of processing pipelines: optical character recognition (OCR) combined with rules, layout-aware machine learning, OCR-assisted large language models (LLMs) and multimodal models working directly on page images. They are compared in terms of extraction accuracy and reliability, robustness to layout and schema changes, and implementation requirements. On this basis, a qualitative decision framework is proposed that links document structure, layout variability, availability of labelled data, accuracy requirements and operational limitations to a justified starting point for testing. The reviewed evidence indicates that effectiveness depends on the task and on the configuration of the whole pipeline, and that OCR accuracy alone does not determine extraction quality. As no approach is universally best, task-specific evaluation with explicit validation is recommended.
    Keywords intelligent document processing, information extraction, OCR, layout-aware models, multimodal large language models
    Full article: PDF icon 39_45_wsakowska_csmm_23_2026.pdf