Research Publications

Abstract:
Software repositories evolve through continuous sequences of changes, ranging from minor edits to major refactorings and feature additions, yet distinguishing genuinely innovative modifications from routine updates remains challenging. This study introduces a version-aware novelty metric that quantifies code innovation by integrating structural dissimilarity with pattern rarity, formalizing the principle that each version's contribution is proportional to the magnitude of its changes. Structural divergence is measured using normalized edit distance against all previously observed patterns, while pattern rarity emphasizes uncommon code structures. The metric produces bounded, interpretable values reflecting both uniqueness and structural innovation. We define three version-aware measures: global novelty, capturing a function's distinctiveness relative to repository history; initial drift, quantifying cumulative deviation from the first implementation; and local change, representing immediate modifications between consecutive versions. Together, these measures form an evolution trace vector, enabling precise monitoring of incremental updates, major refactorings, and innovation peaks. The methodology is language-agnostic, computationally efficient, compatible with standard Git workflows, and demonstrates sensitivity to both substantial rewrites and fine-grained edits, providing actionable insights for automated innovation tracking, risk-aware code review, and AI-assisted development.

Abstract:
The Open University of Sri Lanka (OUSL), serving over 40,000 distance learners, faces persistent challenges in academic planning and student support due to its decentralized and flexible learning model. Key gaps include difficulty visualizing degree progress, lack of personalized study plans, challenges in selecting next subjects based on prerequisites, limited alerts for registrations, and reliance on informal peer support. Existing Learning Management Systems (LMS) are primarily content-centric and activity-based, lacking integrated academic advisory and proactive support features. This study identifies requirements for designing a Smart Academic Assistant, an intelligent support system integrated with OUSL’s Moodle platform, Open University Learning Management System (OULMS). Using a Design Science Research (DSR) methodology supported by Agile development and Scrum, the study followed iterative design, development, and evaluation phases. Quantitative data were collected from 133 undergraduate students across six faculties, complemented by structured interviews with 50 purposively selected students from Level 3-5. Findings revealed that, although academic planning is formally managed by departments, students often struggle with understanding prerequisites, calculating GPA, and navigating registration processes, even after attending pre-induction and induction programmes. Over 97% of respondents expressed interest in an intelligent academic planning tool. The Smart Academic Assistant design incorporates a personalized Academic Scheduler, registration alerts, GPA calculator, degree progress visualization, and an AI-powered chatbot. Following a User-Centred Design (UCD) approach, the system addresses critical challenges: 26.3% of students missed registration requirements due to lack of information, and over 65% desired tools to visualize progress and receive personalized study plans. The application is expected to enhance student autonomy in academic decision-making, reduce reliance on peers, and alleviate administrative burdens. These findings highlight strong demand for improved academic support tools, informing potential enhancements to OUSL’s digital infrastructure. While the study supports the proposed design, the system’s effectiveness requires formal testing with students. Future recommendations include full-scale implementation, SIS-OMIS integration, predictive analytics for at-risk students, and collaboration with other distance learning institutions.

Abstract:
This study introduces SinglishGPT, a transformer-based adaptation designed to enhance real-time translation and information retrieval for Singlish chatbots. Singlish, Sinhala rendered in Latin script, has become a dominant mode of digital communication in Sri Lanka but poses significant challenges for natural language processing due to code switching, orthographic variability, and differences in syntax compared to English. Existing tools, such as transliteration keyboards or generic multilingual models, offer partial solutions but fail to adequately address short, mixed-language queries that dominate real-world usage. This creates a gap in both practical chatbot deployment and theoretical approaches for low-resource, transliterated, and code-mixed languages. The proposed framework combines lightweight normalisation of transliteration variants with parameter-efficient fine-tuning of a multilingual transformer. A compact corpus (~50k tokens) of user-generated Singlish text was curated, pre-processed, and used to adapt the model through adapter-based updates. The pipeline preserves linguistic diversity while stabilising input, enabling the system to learn Sinhala SOV structures, variant spellings, and frequent English insertions. Evaluation used both automatic metrics (BLEU, METEOR, chrF, perplexity) and human judgements from native speakers on fluency, cultural appropriateness, and comprehension. Results demonstrate that SinglishGPT significantly outperforms baselines. It achieved BLEU 0.68, METEOR 0.72, and chrF 0.76, with perplexity reduced from 18.2 to 12.4 compared to a multilingual mT5 baseline. Human ratings averaged 4.2/5, with strong inter-annotator agreement, confirming improved handling of common mixed-language requests. Latency averaged ~300 ms, ensuring interactive feasibility even in cost-sensitive environments. Error analysis highlights residual difficulties with rare abbreviations and emerging slang, pointing to active learning and corpus expansion as future directions. The study concludes that targeted adaptation, combining variant normalisation with parameter-efficient fine-tuning, yields practical and scalable improvements in Singlish chatbot performance. Beyond Sri Lanka, the framework provides a transferable template for other low-resource, transliterated languages where code switching and orthographic variation dominate.

Abstract:
As software systems grow in complexity, integration testing remains a critical yet challenging phase of the software development life cycle, often hindered by error propagation across module boundaries and the limitations of manual debugging techniques. This research introduces an approach that leverages Large Language Models (LLMs) such as OpenAI’s Codex and Meta’s CodeLlama to proactively detect integration-level errors and recommend automated code modifications. The study proposes an intelligent system architecture that incorporates LLMs into the integration testing workflow, comprising modules for static code analysis, fault prediction, and real-time developer guidance through an IDE-integrated interface. The study adopts a mixed-methods approach, sourcing data from real-world, open-source repositories including Microsoft’s Inferred Bugs, the BEARS benchmark, and curated datasets like bugrepo. Empirical evaluation across five enterprise-scale software projects revealed that the LLM-based system achieved a precision of 83% and a recall of 78% in detecting integration faults, with a top-3 accuracy of 69% for code repair suggestions. Developer feedback highlighted a 35% reduction in debugging time and high usability, although challenges remain regarding false positives, domain-specific generalization, and model interpretability. The study addresses a critical gap in the current software engineering landscape by transitioning from reactive to proactive error detection, aligning with contemporary DevOps and Continuous Integration practices. The findings underscore the potential of LLMs to augment software quality assurance processes, reduce development overhead, and accelerate release cycles. Future work will explore domain-specific fine-tuning, adaptive learning strategies, and expansion into system-level testing. The study contributes to advancing intelligent automation in software testing, positioning LLMs as viable assistants in complex, modular development environments.