Authors - Dzhansel Abtula, Stanka Hadzhikoleva, Emil Hadzhikolev, Iliana Ivanova, Elitsa Dubarova Abstract - The present study examines the potential of generative language models as tools for discourse analysis of war-related language in media texts. The study is based on a corpus of texts processed by three generative AI models using an identical prompt that defines a multi-stage analytical procedure. This procedure includes the extraction of war-related lexical units, semantic classification, functional analysis of evaluative and ideological features, compilation of a thematic glossary, analysis of metaphors, identification of discursive strategies, and genre determination of the texts. The generated analyses are evaluated through a structured questionnaire based on a five-point Likert scale, completed by an expert with an academic background. The study aims to systematically assess and compare the quality of discourse analyses produced by different language models under controlled conditions. The results indicate that all models generate structurally coherent and terminologically consistent analyses, but differ significantly in interpretative depth and contextual sensitivity. These findings support the need for a hybrid approach that combines automated analysis with human expertise to ensure accurate interpretation of implicit meanings, ideological nuances, and context-dependent discourse features.
Authors - Makomborero Murwira, Dane Brown Abstract - The proliferation of realistic synthetic speech poses significant threats to information integrity and public safety through financial fraud and misinformation campaigns. While traditional countermeasures based on Gaussian Mixture Models have proven effective against earlygeneration deepfakes, they struggle to generalise to sophisticated attacks produced by contemporary neural synthesis. This paper investigates enhanced audio deepfake detection through Light Convolutional Neural Networks (LCNNs) combined with Linear Frequency Cepstral Coefficients (LFCCs) and advanced training strategies. This study systematically evaluates the impact of margin-based loss functions (Cosface and A-Softmax) and FreqAugment data augmentation on model robustness and generalisation capability. Validated on the ASVspoof 2019 Logical Access dataset, the optimised LCNN model achieves an Equal Error Rate (EER) of 5.40% on the evaluation set containing thirteen unseen attack types, representing a 33% relative improvement over the baseline LFCCGMM countermeasure (8.09% EER). The combination of Cosface loss with FreqAugment demonstrates superior performance compared to ASoftmax configurations, reducing false negatives substantially. Per-attack analysis reveals robust performance across diverse spoofing techniques, though vulnerabilities to advanced neural waveform manipulation methods remain. The proposed framework provides a practical, deployable solution for audio deepfake detection in real-world security-critical applications.
Authors - Ida Bagus Dwipayana , Naniek Utami Handayani, Singgih Saptadi Abstract - The global construction sector is under increasing pressure to shift towards sustainable operational models, particularly in heavy materials processing like asphalt production. This study provides empirical, case-studybased insights regarding sustainable procurement practices at the Asphalt Mixing Plant (AMP), run by PT. Jasamarga Tollroad Maintenance in Karawang, West Java, Indonesia. Utilizing six semesters of the Environmental Management Monitoring Effort (UKL-UPL) data from Semester 1 (S1) 2021 to S1 2025, this research assesses the environmental performance of three integrated innovations: (1) a gas-fired burner system substituting for fuel oil, (2) multi-layered noise reduction technologies, and (3) collection and water vapor-based treatment of stone dust to optimize novelty soil fertility systems. Stack emissions analysis found significantly lower levels of CO at 184 mg/m3 and NOx at 212 mg/m3 than Indonesian regulation enables KEP-13/MENLH/3/1995, while ambient air was consistently compliant during monitoring over nine periods. Outdoor sound levels in the upwind and downwind locations were 65.6 dB (A) and 56.8 dB (A), respectively, both below the exposure limit of 70dB(A) specified by SK Men LH KEP-48/MenLH/11/1996, Sensitivity analysis between four fuel-technology scenarios shows a combination approach achieving a 61.9% CO emission index reduction, 55.4% NOX index reduction, and SO2 index is reduced by 78%, relative to diesel fuel basis. This study finds a significant research gap with respect to operationalizable sustainable procurement criteria combining air quality, noise, and waste valorization in an integrated framework at the AMP operational level. The results offer theoretically underpinned and empirically verified procurement criteria, which are directly applicable to the international construction industry, thus advancing green supply chain management, principles of circular economy as well as pathways towards net-zero construction.
Authors - Cossi Blaise Avoussoukpo, Amara Camara, Babou Dione Abstract - Digital scholarly infrastructures increasingly shape academic visibility by supporting the creation, organisation, dissemination, and discovery of research outputs. Existing studies have advanced important dimensions of this domain, including bibliometric evaluation, persistent identiers, metadata standards, and indexing systems. However, they primarily examine these components independently and provide only a limited understanding of how their interactions inuence visibility in contemporary research ecosystems. This paper introduces the Academic Visibility Pipeline Model (AVPM). This conceptual framework conceptualises academic visibility as a sequence of interconnected stages linking research production, metadata structuring, scholarly indexing, identity and aliation aggregation, dissemination, and discovery. By integrating socio-technical coordination, interoperability, and cascading failure, the AVPM explains how visibility emerges and how upstream constraints inuence downstream outcomes. An illustrative application to the Guinea academic visibility ecosystem demonstrates the explanatory value of the framework by identifying structural bottlenecks and tracing their eects across the visibility pipeline. The proposed framework extends conventional publication- and citation-centred perspectives and provides a process-oriented foundation for institutional strategy, research governance, and digital transformation in emerging research ecosystems.
Authors - Duy Nguyen Ngoc, Hiep Nghia Phan Abstract - Large Language Models (LLMs) have recently shown strong capabilities in automatic text summarization. However, applying these models to lowresource languages such as Vietnamese remains challenging due to limited training resources and language-specific characteristics. In this work, we ex-amine whether prompt optimization can improve Vietnamese summarization quality without modifying model parameters. This paper introduces a Vietnamese-aware Prompt Optimization Framework that refines prompt instructions by combining task-specific guidance, role-based prompting, linguistic con-straints, and iterative feedback. The generated summaries are assessed using both automatic evaluation metrics and human judgments to examine the ef-fectiveness of different prompt designs. We evaluate the proposed approach on a benchmark Vietnamese multi-aspect opinion dataset using several commercial and open-source LLMs, including GPT-4, Claude 3, Gemini 1.5, and PhoGPT, and compare their performance with Vietnamese pre-trained summarization models such as ViT5 and BARTpho. Our experiments show that refining prompts consistently improves summary quality across the evaluated models. In particular, GPT-4 with the optimized prompt achieves an 8.7% increase in ROUGE-L and receives higher human evaluation scores for fluency and factual consistency than the standard prompting setting.
Authors - Tien-Dao Luu, Viet Truong Xuan, Doan Hoang Phuc Nguyen, Nghi Huynh Quang, Do Chau Giang Nguyen, Nghia Nguyen Khoi, Huu Hoa Nguyen Abstract - The Mekong Delta agroecosystem faces compounding climate stressors and acute data fragmentation. While agricultural data exists online, it remains largely unstructured and unverified. This study introduces WikiCrop-AI, an integrated data-processing and machinelearning framework designed to consolidate heterogeneous agronomic information. The architecture comprises three interconnected modules: an Agricultural Notebook utilizing a Retrieval-Augmented Generation pipeline for multi-format data ingestion, a browser-based computational environment (WikiLab) for reproducible workflows, and a client-side analytics module for multivariate clustering. Evaluation of the ingestion pipeline yielded a String Similarity Score of 0.99 for structured text extraction. Furthermore, generative outputs assessed via the Automatic LLMs Citation Evaluation framework achieved average scores of 0.94 for both Citation Recall and Precision, alongside a Claim Recall of 0.92, indicating reliable knowledge synthesis under the tested conditions. A hierarchical clustering case study on 22 soybean cultivars further illustrates the platform’s utility in supporting non-programming local agronomists. Ultimately, WikiCrop-AI provides a decentralized infrastructure to translate scattered digital resources into actionable, verified agricultural intelligence. The complete open-source code for the WikiCrop-AI ecosystem is publicly accessible on GitHub at: https://github.com/mekonglab-vn/ WikicropAI.
