UCSC Digital Library Collection:

Hate Speech Detection in Sinhala and Romanized Sinhala Mixed with Emojis using Hybrid CNN-RNN Model

2025-06-24T00:00:00Z

Title: Hate Speech Detection in Sinhala and Romanized Sinhala Mixed with Emojis using Hybrid CNN-RNN Model Authors: Waidyarathne, H K Abstract: ABSTRACT The paper addresses the critical issue of the hate speech detection in online Sinhala and Romanized sinhala content associated with emojis. The widespread of online hateful content is particularly significant in low-resource languages such as Sinhala, where natural language processing tools and resources are limited. The study examined the development and evaluation of a novel deep learning approach to effectively identify the words of hatred in the Sinhala language and the Romanized Sinhala text, and complicated the inclusion of emojis, which are prevalent in the online communications of the Sinhala language. In order to address this challenge, we propose a combined model of the Convolutional Neural Network - Bidirectional Long Short-Term Memory (CNN-BiLSTM) as our primary architecture. This model is designed to use the strengths of both CNNs to capture local n-gram patterns and BiLSTMs to understand the sequential context within the Sinhala text. CNN extracts local signs of hate speech, while BiLSTM captures long-term dependencies and contextual information in Sinhala comments. An Embedding layer, shared by both branches, learns dense vector representations for Sinhala words and emojis optimized for the task of detecting offensive speeches. In order to compare, we also implemented and evaluated independent CNN, BiLSTM, LSTM, Support Vector Machine (SVM), Random Forest, Logistic Regression, and Naive Bayes models. The results show that our CNN-BiLSTM combined model is superior to the comparable models. The CNN-BiLSTM model achieved a high accuracy of 0.9251 and a balanced F1 score of 0.92 for the “Offensive” class, indicating the robustness of detecting hatred speech while minimizing both positive and negative false information. The analysis of the confusion matrix further revealed a balanced distribution of errors, highlighting the effectiveness of the model in the treatment of offensive and non-offensive Sinhala content. Although simple models such as SVM and Random Forest showed competitive accuracy, their precision and F1 points were generally lower, especially in the “Offensive” class. Neural network architectures, in particular CNN-BiLSTM and CNN, have consistently surpassed traditional machine learning methods. Keywords: Hate Speech Detection, Sinhala Language, Low-Resource Language, CNNBiLSTM, Deep Learning, Natural Language Processing, Romanized Sinhala, Emojis

A Comparative Analysis of Provisioned Concurrency Types for Cloud-Based Applications

2025-07-12T00:00:00Z

Title: A Comparative Analysis of Provisioned Concurrency Types for Cloud-Based Applications Authors: Rathnayaka, M N J Abstract: Abstract The increasing adoption of serverless computing, particularly AWS Lambda, has introduced both performance benefits and challenges. A major concern is the cold start latency, which affects response times and scalability, particularly for time-sensitive applications. To mitigate this issue, AWS introduced provisioned concurrency, which ensures pre-initialized function instances to reduce startup delays. However, selecting the appropriate concurrency model static or dynamic (scheduled or auto-scaling) remains a challenge due to the trade-offs between performance and cost. This dissertation presents a comparative analysis of provisioned concurrency types in AWS Lambda, evaluating their impact on various workload domains, including web applications, IoT backend services, and data processing systems. Real-world and simulated workloads were analyzed using statistical techniques, such as Mahalanobis distance calculations, to assess performance variations. A simple rule-based guide was developed to assist in selecting the most suitable concurrency type based on application needs, balancing cost efficiency and performance. For IoT workloads, scheduled provisioned concurrency (Mahalanobis distance: 10.80; cost: $0.0888/1000 invocations) achieved optimal balance, reducing latency by 8.7% over static provisioning. E-commerce applications benefited most from static provisioning (Mahalanobis distance: 9.00; cost: $0.0054/1000 invocations), ensuring low latency during traffic spikes. Data processing systems favored scheduled concurrency (Mahalanobis distance: 3.009; cost: $3.576/1000 invocations) for predictable batch jobs. Cost-Distance Ratio (CDR) analysis further validated these recommendations, prioritizing strategies with CDR < 1. The study provides empirical guidelines to optimize AWS Lambda deployments, recommending scheduled concurrency for IoT, static for e-commerce, and scheduled for data processing, ensuring cost-effective performance across domains.

Enhancing Tamil Text Readability for Dyslexic Readers Using Generative AI Based Text Simplification

2025-06-30T00:00:00Z

Title: Enhancing Tamil Text Readability for Dyslexic Readers Using Generative AI Based Text Simplification Authors: Rajendran, K Abstract: ABSTRACT Dyslexia is a commonly recognized reading disorder impacting reading fluency, comprehension and cognitive processing. While extensive research has been conducted on assistive technologies for dyslexic readers, most solutions are primarily designed for high resource languages such as English. Consequently, accessibility tools for Tamil readers remain scarce. The agglutinative morphological complexity of the Tamil language, along with its extensive use of ligatures and visually similar characters, further exacerbates reading difficulties for dyslexic individuals. This study addresses this research gap by proposing a Large Language Model based text simplification approach aimed at improving text readability while maintaining semantic coherence. Adopting a Design Science Research methodology, this study integrates few-shot learning, prompt engineering, and rule-based text simplification techniques to enhance Tamil text accessibility. A custom curated dataset incorporating Tamil text with predefined simplification rules was developed and used to evaluate the performance of various LLMs, including BART, T5 variants and LLaMA variants. Experimental evaluations were conducted using both qualitative feedback and quantitative metrics to assess the effectiveness of these models. The findings indicate that prompt engineering alone resulted in limited improvements, whereas fine-tuning domain-specific LLMs trained on Tamil linguistic rules demonstrated greater efficacy in generating simplified text. However, challenges persist, particularly in adapting LLMs to Tamil’s intricate linguistic structure and ensuring that readability enhancements align with dyslexia-friendly formatting standards. This research contributes to the broader field of Natural Language Processing for low-resource languages by demonstrating the feasibility of LLM driven Tamil text simplification. Future research directions may include fine-tuning larger Tamil-specific LLMs, integrating reinforcement learning with human feedback, and deploying real-time text simplification models for assistive applications. By addressing the accessibility needs of Tamil-speaking dyslexic readers, this study paves the way for inclusive AI-driven text simplification solutions, ultimately improving the readability and comprehensibility of complex textual content.

Hybrid Deep Learning System for Nail Disease Diagnosis Combining Synthetic Data, Explainable AI and Multi-Model Architectures

2025-06-25T00:00:00Z

Title: Hybrid Deep Learning System for Nail Disease Diagnosis Combining Synthetic Data, Explainable AI and Multi-Model Architectures Authors: Attanayake, T. M. D. R. M Abstract: ABSTRACT Nail abnormalities serve as crucial indicators of a wide range of systemic, dermatological and nutritional disorders, necessitating early and precise detection for effective medical intervention. However, conventional diagnostic methods rely on subjective visual assessments by healthcare professionals which can lead to inconsistencies and delays in diagnosis. This research presents an AI-powered Nail Abnormality Detection System which has been designed to automate the identification and classification of five common nail abnormalities which are clubbing, spoon nails (koilonychia), black nails, splinter hemorrhages and nail pitting alongside normal nails. By integrating advanced deep learning techniques, this study addresses key challenges in medical image analysis including data scarcity, model interpretability and real-time diagnostic efficiency. The proposed system utilizes a hybrid deep learning framework combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to enhance feature extraction by improving both classification accuracy and generalization. Given the limited availability of high-quality annotated datasets, this research also explores synthetic data generation using diffusion models to augment training data while ensuring balanced representation of various nail conditions. In addition, Explainable AI (XAI) techniques such as Grad-CAM, Ablation-CAM and SIDU, are employed to provide visual interpretability, enhancing model transparency and enabling medical professionals to validate AI-driven predictions. Furthermore, YOLO-based segmentation is integrated to precisely localize affected regions by facilitating severity assessment and aiding in clinical decision-making. Comprehensive experimental evaluations demonstrate that the system achieves great classification accuracy, high precision and recall across multiple datasets while outperforming traditional diagnostic approaches. The findings establish that the combination of deep learning and computer vision techniques significantly enhances diagnostic reliability, offering a scalable, accessible, and cost-effective solution for early disease detection. The study not only advances the field of AI-driven dermatological diagnostics but also lays the groundwork for future research in AI-assisted healthcare applications. With its ability to provide rapid, consistent and interpretable results, this system has the potential to revolutionize telemedicine, point-of-care screening and large-scale population health monitoring. The pipeline begins with a Nail-Filter CNN that screens photographs for relevance, achieving 98 % overall accuracy, 100 % recall for nail images and 100 % precision for non-nail images (F1 = 0.98), thus eliminating most irrelevant inputs before further processing. A hybrid