University of Ibadan Journal of Science and Logics in ICT Research

A Natural Language Processing (NLP) Model for Metaphor Detection and Interpretation: A Case Study of Use of English Passages in UTME

2026-06-13T14:52:24+00:00

Metaphors play a fundamental role in language comprehension. They convey abstract concepts through vivid imagery and analogy, and can help students to understand written texts. In Natural Language Processing (NLP), the accurate detection and interpretation of metaphors pose significant challenges because of their complexity and contextual variability. This study developed an NLP model for metaphor detection and interpretation, using sentences from the ‘Use of English’ passages in Unified Tertiary Matriculation Examination (UTME) past questions as a case study. The approach involved training a transformer-based RoBERTa model on the Vrije Universiteit Amsterdam metaphor corpus (VUA-20), and fine-tuning it on a dataset built from UTME comprehension passages. Contextual embeddings and Word Sense Disambiguation (WSD) were used to interpret metaphorical meanings. The results showed promising performance in metaphor detection, with precision, recall, F1 score and accuracy values that indicated the effectiveness of the model on both datasets. The interpretation step also produced literal meanings for detected metaphors, which can aid language comprehension in an educational context. The study confirmed that transformer-based NLP models can be adapted to specific domains for metaphor detection and analysis.

A Smart Contract–Enabled Private Ethereum Blockchain for Secure Academic Transcript Verification

2026-06-13T13:07:02+00:00

Forgery and delayed validation of transcripts by Nigerian universities had necessitated the development of reliable systems. Inadequacies in manual, centralized approaches had been characterized by inefficiencies, high chances of corruption, and forgery of academic documents. In this paper, an Ethereum blockchain solution for the Management of Computerized Instructional Universities (MCIU) was presented to guarantee authenticity, reliability, and confidentiality of student transcripts. This proposed solution incorporated Solidity smart contracts on a private Ethereum network for handling transactions related to the issuance, reissuance, and validation of transcripts where transcript files were stored on the decentralized file storage (Swarm), while only hashes of the transcript files were recorded on the blockchain. React and Node.js technologies were employed in designing both client and server sides of the system. Authorized personnel were able to issue digital verifiable transcripts, and third parties were able to validate issued transcripts within seconds through a secure coding process.

Abstract The Student Industrial Work Experience Scheme (SIWES) is a crucial component of technical and vocational education in Nigerian institutions, designed to bridge the gap between classroom learning and industry practice. However, traditional SIW

2026-06-13T11:05:41+00:00

Abstract The Student Industrial Work Experience Scheme (SIWES) is a crucial component of technical and vocational education in Nigerian institutions, designed to bridge the gap between classroom learning and industry practice. However, traditional SIWES logbooks, maintained in paper format, pose challenges such as limited accessibility, data loss, and difficulty in leveraging documented experiences for analytics and grading. This study presents the design and development of a Cloud-Based Logbook-as-a-Service (LaaS) for SIWES, enabling real-time documentation, retrieval, and analysis of students’ industrial experiences. The system has an integrated assessment module to grade student performance based on logged activities and feedback from industry supervisors. This study provides a system that fosters digital transformation on SIWES documentation and grading processes by providing a centralized, scalable, and accessible platform, this innovation enhances the efficiency of SIWES evaluation and fosters data-driven decision-making for academic and industry stakeholders. To assess the usability of the developed platform, the system was evaluated based on user experience and task completion time. The overall system performance, measured using these metrics, indicated positive outcomes.

Theoretical Analysis of the Universal Approximation Properties of GELU in Neural Networks

2026-06-11T13:15:01+00:00

The choice of activation function is critical to a neural network’s expressive power. The Rectified Linear Unit (ReLU) became a widely adopted standard due to its computational efficiency and effectiveness in mitigating vanishing gradients. However, ReLU also possesses well-known theoretical limitations, including non-differentiability at zero and the "Dying ReLU" problem, which can impede training. As an alternative, the smooth ????? Gaussian Error Linear Unit (GELU) has seen increasing adoption in state of-the-art models. This paper provides a rigorous theoretical analysis of GELU’s universal approximation properties. We formally prove that GELU satisfies the necessary and sufficient conditions of the Universal Approximation Theorem (UAT) by demonstrating that its ????? smoothness ensures its membership in the required function class ?, and that its non-terminating Taylor series expansion proves its essential non-polynomial nature. To support this theoretical analysis, we present a series of targeted empirical validations that visually and quantitatively demonstrate the practical consequences of these properties. Our experiments confirm that GELU’s smoothness provides a tangible advantage over ReLU in approximating ????? functions, especially in deep neural networks; its non-zero negative gradient prevents the neuron death seen in ReLU; and its unbounded nature is superior to Tanh for modeling non-saturating functions. This work provides a complete theoretical explanation for GELU’s power as a universal approximator, bridging the abstract UAT framework with the function’s specific mathematical properties.

Software Security Vulnerability Prediction Modelling for C/C++ Systems

2026-06-11T09:03:12+00:00

This study focused on developing realistic software security Vulnerability Prediction Models (VPMs) for C/C++ systems. The aim is to mitigate security vulnerabilities and prevent exploitation in C/C++ projects by identifying vulnerable source files for patching before deployment. The study addressed the limitations of existing software VPMs, such as low accuracy, poor traceability of vulnerabilities, dataset imbalance, and the use of irrelevant metrics. The research used relevant security-related metrics as features and addressed the dataset imbalance issue by oversampling. Genetic algorithm was modified to overcome local optima problem and in turn used to optimize the correlation values of the metrics and improved the performance of random forest classifier. The study also highlighted that oversampling improved predictability and feature elimination mitigated overfiting. The developed software VPMs exhibited high performance in cross-project predictions, with recall, precision, and f-measure exceeding 80%, surpassing most performance reported in the literature. The software VPMs enable easy traceability of vulnerable components. Therefore, the study recommended the adoption of these software VPMs by quality assurance teams in software development companies to predict vulnerable files to patch before deployment. Additionally, the primary dataset used in the study is recommended as a benchmark for software VPMs researchers.

A Prediction Model for Cardiovascular Health Risk from Air Quality Index of Pollution laden Environment

2026-06-13T15:09:39+00:00

The cardiovascular health concerns are triggered by various causative agents, of which environmental inhaled pollutants such as CO, NO, NO2, PM10, PM2.5, O3 and SO2 from is an important agents. The introduction of air pollutants is caused by human activities that introduce contaminants into the air. In Nigerian, people living in pollution laden environments are unknowingly exposed to these risks such as Asthma, cough, lung cancer etc. However, there is paucity of information on the health risk impacts on the people living in pollution laden environments. This is due to lack of predictive model to reveal the associated risk to enhance early detection and prevention. One of the methods to evaluate and predict the pollutant is the use of Air Quality Index (AQI) dataset. The quality of AQI data of an environment is a pointer to the degree of pollution and the health risk of the inhabitants. Existing predictive techniques such as Probability and Statistics model used to predict AQI were very complex with some level of uncertainty which necessitate an alternative approach for better accuracy. A Machine Learning (ML) approach combined with an associative decision rule was used to predict the air quality and to identify areas predominates with toxic air quality. Two datasets; open and locally sourced were used, data pre-processed and engineered implementation were done using python coding. The prediction models; Support Vector Classifier (SVC) and Random Forest Classifier (RFC) were employed. The performances of the models were evaluated using classification reports and confusion matrix metrics. The RFC gave an accuracy level of 99% and SVC an accuracy level of 83%. This results show that AQI predictions obtained through RFC is better in accuracy when compared with SVC.

Modified Poll Search Differential Evolution Algorithm for Global Optimisation

2026-06-10T08:07:42+00:00

This paper introduces MPS-DE, an innovative variant of Differential Evolution (DE) that integrates the Modified Poll Search (MPS) technique with DE. The incorporation of MPS seeks to enhance the efficiency and effectiveness of the DE algorithm. A comprehensive empirical comparison was conducted among MPS-DE, RCGA-P, SRCGA, and DE/Best/1/Bin on a suite of twenty global optimisation benchmark functions. The results show that MPS-DE demonstrates superior performance compared to DE/Best/1/Bin and SRCGA, although RCGA-P outperforms MPS-DE in certain benchmarks, indicating the impressive and potential search capabilities of MPS in evolutionary algorithms.

Cargo Revenue Prediction Model Using Machine Learning Approach

2026-06-13T13:39:04+00:00

Abstract This study develops a machine learning (ML) predictive model tailored to Nigeria's cargo ecosystem, aiming to enhance revenue forecasting, strategic planning, and operational efficiency. Data encompassing 1,133 records from Nigerian logistics firms (detailing variables such as cargo weight, shipping rates, and transaction dates) was preprocessed using one-hot encoding, normalization, and median imputation. Four primary regression models (Decision Tree [DTR], Random Forest [RFR], Gradient Boosting [GBR], and a Stacked Adaptive Multi-Input Regression Algorithm [SAMIRA]) were deployed via Google Colab. Exploratory Data Analysis (EDA) revealed right-skewed revenue distributions and seasonal operational peaks in January, August, and September. Model evaluation demonstrated that GBR outperformed the others, achieving an R² of 0.9989, Mean Squared Error (MSE) of ?2.74 billion, Root Mean Squared Error (RMSE) of ?52,350.29, and Mean Absolute Error (MAE) of ?11,213.09. This superior performance was validated through 10-fold cross-validation (mean R² = 0.9969) and further visualized via a normalized error heatmap. Subsequently, the optimal model was prototyped into a Kotlin based Android application for real-time forecasting. The findings demonstrate that GBR can achieve >99% forecasting accuracy, presenting a robust alternative to traditional methods and offering actionable insights for dynamic pricing and resource optimization in emerging markets.

Development of a Deep Learning Model for the classification of Alzheimer’s Disease from Magnetic Resonance Imaging

2026-06-13T11:41:47+00:00

Alzheimer’s disease (AD) is a disorder in which the nervous system slowly declines progressively which affects reasoning, forgetting, balancing, daily activities and memory. The early recognition and diagnosis help to manage and treat the disease effectively. Magnetic Resonance Imaging (MRI) especially 3D scan brain imaging which provides detail structural format and information that can aid in recognizing anomalies linked to the various stages of Alzheimer’s disease (AD). The problem face with manual interpretation of MRI scans is enormous in terms of accuracy and time consuming with clinical experts. In this study, we propose a deep learning approach for the multi classification of Alzheimer’s disease from 3D MRI images. The framework uses Convolutional Neural Networks (CNNs) for developing intelligent model for effective 3D image analysis and interpretation. To enhance classification performance, the extracted region of interest is modified with deep learning classifiers including Efficient Net, SE-ResNet and Dense Net. These architectures improve feature representation, enhance efficiency and improving learning capability of the framework. The results shows that the model achieves accuracy of 83% and precision of 82%, which indicates strong performance. The recall and F1-sore display a balance of 81 % across ford in a distinct phase in the progression of Alzheimer’s disease multi class classification. This model will assist the clinicians and radiologist in early interpretation, detection, diagnosis and monitoring of Alzheimer’s disease progression.

Smart Leak Detection in Water Distribution Networks Using Hybrid Deep Learning Models

2026-06-13T09:56:41+00:00

Water leakage in distribution networks poses significant challenges due to aging infrastructure, rising demand, and the limitations of conventional detection methods, resulting in substantial water loss and increased operational costs. This study proposes a hybrid deep learning approach that combines a Convolutional Neural Network (CNN) and a Support Vector Machine (SVM) for efficient leakage detection in water distribution networks. The CNN is utilized to automatically extract high-level features from multivariate sensor data, while the SVM performs robust classification to improve generalization and decision accuracy. A real-world water network dataset containing pressure, flow rate, and velocity measurements was used for model development. After data cleaning, feature selection, and Min–Max normalization, the dataset was split into training and testing sets. Model performance was evaluated using accuracy, precision, recall, F1-score, and ROC-AUC metrics. Experimental results indicate that the proposed CNN–SVM hybrid model attains 95% accuracy and a ROC-AUC score of 97%, outperforming CNN models. The results confirm that integrating deep feature extraction with machine learning classification enhances leakage detection reliability. This approach provides a scalable and effective solution for real-time monitoring of water distribution networks and contributes to reducing non-revenue water and improving sustainable water resource management.

Attention-Based LSTM Model for Malaria Severity Prediction in Bayelsa State using Clinical, Environmental and Geospatial Data

2026-06-11T11:08:50+00:00

Malaria remains a critical public health concern in Nigeria, with Bayelsa State experiencing persistent transmission due to its tropical climate, riverine geography, and seasonal flooding. Early identification of malaria severity is essential for effective clinical management, reduction of complications, and optimal allocation of limited healthcare resources. This study presents an attention-based Long Short-Term Memory (LSTM) deep learning model for predicting malaria severity levels, categorized as low, moderate, and high using integrated clinical, environmental, temporal, and geospatial data collected within Bayelsa State. The dataset comprised patient demographic attributes, clinical indicators such as body temperature, environmental variables including rainfall and climate temperature, and geospatial information at the local government area (LGA) level. Rigorous data preprocessing, feature engineering, and data leakage prevention techniques were employed to enhance model reliability. Class imbalance was addressed using the Synthetic Minority Over sampling Technique (SMOTE) and class-weighted training. Experimental evaluation using multiple performance metrics demonstrated that the proposed attention-based LSTM model achieved strong and balanced predictive performance across all severity classes. The results underscore the effectiveness of deep learning with attention mechanisms for malaria severity prediction and highlight its potential application as a clinical decision support tool in malaria-endemic regions such as Bayelsa State.

A Hybridized Data Mining Technique for Enhanced Network Intrusion Detection Performance

2026-06-10T14:16:05+00:00

Conventional intrusion detection systems (IDS) are no longer efficient enough to recognize newly designed cyber-attacks because of increasing complexities and amount of network traffic data. A more effective approach for Data Mining (DM) is required for cybersecurity applications, although individual Data Mining strategies were sufficient for intrusion detection systems previously. In order to enhance precision and malleability for network intrusion detection systems (NIDS), this paper proposes a hybrid strategy for data mining using “CIC IDS2017” dataset downloaded from Kaggle. This hybrid approach uses ensemble learning to increase classification efficiency by unifying the “Adaptive Boosting” approach strengths and the “C4.5 Decision Tree” algorithm technique concepts. Data preprocessing techniques, Label Encoding techniques, and classifiers belonging to the “Supervised Classification” category stood as key components of this strategy approach. Its efficiency is assessed using standard metrics. This proposed hybrid strategy approach resulted in near-perfect performance on its testing approach by generating 317,937 “True Positives” values, having “4” “False Positives” values, and having “Accuracy” of 99.9%. The performance of “C4.5 Classifier” approach also resulted in generation of 317,938 “True Positives” values having “5” “False Positives” values having “Precision” “Recall” “and F1-score” measures recorded at 99.9%. “Adaptive Boost” approach resulted in “317,185” “True Positives” values having “287” “False Positives” values having “Accuracy” “Precision” “Recall” “and F1-score” “values at “99.7%” “99.8%” “99.5%”, “and 99.7%”. This enhances development efforts of “intelligent” “cybersecurity” “systems” by applying “Deep” “Learning” “concept” further emphasizing “Data” “Mining” application for “network” “enhancements” to remain efficient

A Systematic Review on Approaches for Evaluating the Effectiveness of the Ponseti Method in Clubfoot Treatment

2026-06-10T09:27:50+00:00

Congenital talipes equinovarus (CTEV), commonly known as clubfoot, remains one of the most prevalent congenital orthopedic deformities affecting newborns worldwide and necessitates effective management strategies. The Ponseti method, comprising serial casting, percutaneous tenotomy, and bracing, continues to serve as the standard for non surgical correction; however, its success is influenced by factors such as the severity of the deformity, timing of intervention, clinician expertise, and patient adherence. This systematic review examines the integration of techniques, including statistical models, machine learning (ML), and Interval Type-3 Fuzzy Logic (IT3FL) methods, alongside ontology-based frameworks that enhance knowledge representation and interoperability for improved clinical decision-making. Drawing insights from 225 studies published between 1963 and 2025, the review identifies a paradigm shift from empirical to data-driven methodologies, with a notable increase in AI-focused research since 2020. Despite these advancements, challenges persist, particularly regarding limited dataset diversity, small sample sizes, and insufficient clinical validation. Future investigations should emphasize large-scale, multi-center collaborations and the development of clinician-oriented intelligent systems to advance personalized and interpretable management of clubfoot

Development of a Multimodal Artificial Intelligence Framework for Forest Fire Prediction

2026-06-13T14:15:09+00:00

This study presents a Multimodal Artificial Intelligence Framework (MAIF) that integrates real-time sensor data and visual imagery to enhance forest fire detection accuracy, responsiveness, and reliability in remote environments. The system combines ensemble classification models such as Random Forest, SVM, KNN, XGBoost and Gradient Boosting with YOLOv8-based image recognition to detect fire risk patterns and visual indicators such as smoke and flame contours. A custom-built Forest Fire Capturing Device (FFCD), equipped with an ESP32 microcontroller and LoRaWiFi, was deployed in Omo forest, Nigeria, to collect heterogeneous environmental data. Visual inputs from ground cameras and drones were fused with sensor-based predictions to minimize false positives and improve generalization. The base classifiers showed performances of 0.98, 0.96, 0.93, 0.98, 0.98 for RF, SVM, KNN, XGBoost and GB, respectively with heterogeneous sensor datasets of 10,334 rows and 13 columns while meta-classifier and YOLOv8 module both achieved 0.98 accuracy, with significantly lower false positive rates compared to single-modality systems. Upon confirmed detection, the system automatically dispatched timestamped fire images via email, enabling rapid situational awareness and emergency response coordination.

Architecting Resilience: A Cloud-Native Neural Risk-Scoring System for Enhanced Campus Security and Streamlined Admissions in Nigerian Universities

2026-06-13T12:08:46+00:00

Nigerian universities face a convergence of pressures that conventional admission procedures were never designed to handle: application volumes that now reach 1.9 million candidates per cycle, growing incidents of campus related violence and organised cultism, and administrative structures whose capacity has not scaled alongside institutional enrolment. The result is a screening process that is simultaneously too slow, too inconsistent, and too shallow to serve its stated purpose of protecting campus communities. This study responds to that gap by designing and specifying the full deployment architecture for a cloud-native system that uses a trained Multi-Layer Perceptron (MLP) to generate quantified pre-admission crime risk scores for individual applicants. The design work reported here builds directly on prior empirical modelling research [3] and extends it across four engineering dimensions: a containerised, cloud-hosted inference infrastructure; a versioned RESTful API layer enabling integration with existing university information systems; a layered data security framework satisfying both the Nigeria Data Protection Regulation (NDPR) and the EU General Data Protection Regulation (GDPR); and a governance structure that keeps human admissions officers firmly in control of final decisions. A three-phase rollout plan is specified to accommodate the financial and technical realities facing most Nigerian higher education institutions, where capital budgets are constrained and IT departments are thinly staffed. Seven tables provide engineering reference data covering screening performance comparisons, MLP configuration parameters, cloud platform trade-offs, deployment considerations, privacy controls, API specifications, and projected operational indicators. Four architectural figures accompany the text. Taken together, the design presented here offers the Nigerian higher education sector a technically rigorous, institutionally calibrated pathway toward evidence-based, consistent, and legally defensible admission screening — one that does not require institutions to trade away ethical accountability in pursuit of efficiency.

A Systematic Review of AI-Powered Assessment and Feedback to Enhance Teaching Effectiveness in Higher Education

2026-06-13T10:28:12+00:00

The growing use of artificial intelligence in higher education assessment offers a real chance to improve teaching effectiveness. However, the practical, ethical, and pedagogical aspects of this shift are not yet well understood. This paper presents a systematic review of 127 studies selected from 699 candidate papers published between 2022 and 2025. The review examines evidence across five themes: feedback personalisation, assessment accuracy, ethical and equity concerns, human-in-the-loop integration, and pedagogical impact. The findings show that AI tools, particularly GPT-4, can deliver personalised, timely, and scalable feedback that improves student engagement and academic outcomes while reducing educator workload. However, AI-generated feedback is often less sensitive to context, empathy, and higher-order thinking tasks, especially in creative and humanities subjects. Hybrid models that combine AI with human oversight are the well-supported approach, as they improve grading accuracy, fairness, and student trust. Issues such as algorithmic bias, data privacy, lack of transparency, and weak governance remain key ethical challenges that need strong institutional responses. This review offers a clear evidence base to guide educators, policymakers, and technologists on how to use AI-enhanced assessment in a responsible and sustainable way.

NETPA-DLA: A Deep Learning–Based Network Packet Analyzer for DDoS Detection

2026-06-11T12:30:42+00:00

In the digital era of internetworked systems, understanding, analysing and filtering network traffic is crucial for maintaining security, optimal performance, conducting diagnostic routines and monitoring. This research developed a novel Deep Learning based Network Packet Analyzer (NETPA-DLA) which utilizes an optimal hyperparameter dynamic technique. An ensemble deep learning approach that integrates Deep Belief Networks (DBN), Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNN), Auto-encoders, Transformers and U-Net for robust and accurate classification of distributed denial-of-service (DDoS) was used. The ensembled model was trained on the CIC-DDoS2019 dataset. The study findings contribute to the continuous refinement and deployment of advanced measures to strengthen digital infrastructure against evolving threats. The experiment on the non-pretrained DBN model proved to be better than the pretrained counterpart for DDoS detection, with an accuracy of 99.72 % and false positives of 37 and false negatives of 13 on the validation dataset, with results for all metrics for the LSTM model at 0.9998, the least being validation specificity at 0.9855. Transformer had the highest accuracy level of 0.9998, closely followed by Autoencoder, which had an accuracy level of 0.9986, and ensemble weighted voting at 0.9984, while the RNN obtained a perfect score of 1.0000 for both Recall and Sensitivity across the three relative weights for each of the models. The study shows that DBN can accurately detect and predict DDoS while maintaining the security of the system and given access to the necessary user of the system without any form of denial.

An Enhancement of African Low-Resource Corpora with NLP IgboT5

2026-06-10T10:11:35+00:00

This paper adopts the Text-to-Text Transfer Transformer (T5) for the Igbo language Natural Language Processing Tasks. IgboT5 enhances the previous digital Igbo Thesaurus through the creation of a high-quality Igbo dataset. The paper fine-tunes a multilingual T5 model and evaluates it on tasks such as definition generation, paraphrasing, translation, and context completion. This paper contributes to the advancement of low resource African languages and opens doors for future Natural Language Processing (NLP) applications.