Conference papers

Permanent URI for this communityhttps://rda.sliit.lk/handle/123456789/4889

Browse

Now showing 1 - 8 of 8

Embargo
A Reinforcement Learning Approach with Human in the Loop to Explainable Insurance Risk Scoring and Intelligent Policy Portfolio Optimization
(Institute of Electrical and Electronics Engineers, 2026-05-29) Gamage, C; Kasthuriarachchi, T; Denuwan, C; Mallawaarachchi, P; Abeywardhana, L; Nawarathne, M
Assessing individual risk accurately and optimizing insurance portfolios in real time remain major challenges due to static actuarial tables, opaque models, and fragmented analytical pipelines. This paper proposes a two-part Explainable AI (XAI) framework addressing both issues. The first component, Artificial Intelligence-driven risk scoring with human-in-the-loop (HIL) weight adjustment, uses a Proximal Policy Optimization (PPO) agent to suggest feature-based changes to an insurer's risk-equation weights. Shapley Additive Explanations(SHAP) attributions and Generative AI reasoning make these changes interpretable, allowing human reviewers to approve modifications that are immediately applied to future customers, creating a self-improving loop. The second component, AI-driven policy optimization, leverages a PPO supported by an XGBoost expense regressor, SHAP/LIME explainability, PPO agent and a Retrieval-Augmented Generation (RAG) layer for rider assignment. Both components share a data backbone of 100,000 anonymized insurance records stored in MongoDB, enabling incremental updates without reprocessing. Experiments show the XGBoost regressor achieves Root Mean Square Error (RMSE) 0.4406 and Mean Absolute Error (MAE) 0.3600, the HIL guided agent increases average episodic reward by 10.3%, and the RAG layer reaches 91.7% rider-assignment accuracy. The framework significantly enhances predictive accuracy, interpretability, regulatory traceability, and portfolio adaptability compared to traditional actuarial and black-box approaches.
Embargo
AI-Assisted Criminal Investigations: Enhancing Testimony Analysis and Case Correlation
(Institute of Electrical and Electronics Engineers Inc., 2025-12-08) Karunarathna K.A.D; Ediriweera E.R.L.C; Fernando N.D.R.N; Abeywardana B.A.I.S; Abeywardhana, L
This paper presents Lexa AI, a holistic AIpowered solution for improving criminal investigative practices through four phases that target key drawbacks found in traditional practices. The first model provides an automated data/information collection stage that accepts legal documents in multiple formats and utilizes a three-stage processing pipeline based on Gemini 2.0 Flash model, which performs Optical Character Recognition (OCR) with a higher level of accuracy and speed compared to alternative approaches. The collected data will be forwarded to the next phase by leveraging dynamic question generation that uses Reinforcement Learning (RL), instantaneous multilingual capabilities, and real-time scoring of relevance. The third phase conducts Multimodal Behavioral and Physiological Analysis (MBPA), which includes facial signals, speech signals, and heart rate signals to create a combined Stress Index as an objective indicator and avoids subjective judgment. Finally, semantic similarity will be measured to correlate incidents, assess risk for victims, and provide explainable predictions
Embargo
An Integrated Approach to Smart Criminal Judgment Analysis in Sri Lankan Courts
(Institute of Electrical and Electronics Engineers, 2026-05-29) Sundaresan, K; Ekanayake, D; Thavachchelvam, N; Sivaanbu, A; Abeywardhana, L; Nawarathne, M
Criminal justice practitioners in Sri Lanka face considerable challenges when accessing and analyzing past court judgments for case preparation, as most research tasks remain manual and time intensive. This study presents an integrated system comprising four interconnected components to address these challenges. A Legal Resource Extractor uses a Hybrid Neuro-Symbolic Architecture with domain-adapted Legal-BERT, fine-tuned on 133,338 legal text chunks, to extract structured legal knowledge from multilingual inputs including voice evidence submitted by lawyers. A Case Analysis and Argument Generation module employs LegalBERT embeddings over 1,601 Court of Appeal judgments with a Nearest Neighbors retrieval model achieving 94% perfect retrieval rate for precedent-based argument generation. The Appeal Outcome Prediction component has been optimized through a Hybrid Feature Engineering pipeline that integrates 1,000 TF-IDF lexical features, 49 traditional legaldomain indicators, and 768-dimensional Legal-BERT embeddings. By utilizing SMOTE and a Calibrated Voting Ensemble, the system predicts appeal outcomes with 79.75% accuracy across three outcome classes. A Public Legal Assistant system provides offline legal information in Sinhala, Tamil, and English through hybrid FAISS based retrieval using 1,146 legal documents and locally hosted language model generation. Together, these components offer practitioners an end-to-end platform for strengthening legal research, argument preparation, and access to justice.
Embargo
Explainable Multilingual Sentiment Analysis for Sinhala, English and Code-Mixed Banking Reviews
(Institute of Electrical and Electronics Engineers Inc., 2025-12-09) Senevirathna, P; Adhikari, N; Navojith, T; Rizvi, A; Kasthurirathna, D; Abeywardhana, L
The growth of multilingual and code-mixed communication in the financial sector presents significant challenges for sentiment analysis, particularly in low-resource languages such as Sinhala. Existing NLP models often underperform due to limited annotated data, domain-specific vocabulary gaps, and a lack of transparency. This paper presents a multilingual and explainable sentiment analysis framework that addresses the challenges of low-resource and code-mixed languages by analyzing banking customer feedback in English, Sinhala, and Sinhala-English code-mixed formats. A dataset of 13,000 aspect-tagged banking reviews spanning five key banking aspects is used for evaluation. The framework classifies comments into positive, neutral, or negative sentiment and provides interpretable explanations using SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model Agnostic Explanations). English sentiment classification is performed using a fine-tuned BERT-base-uncased model, while Sinhala/code-mixed sentiment is handled by a hybrid XLM RoBERTa with a domain-specific lexicon correction approach. The framework achieves 92.3% accuracy and 0.89 F1-score for English, and 88.4% accuracy and 0.84 F1-score for Sinhala/code-mixed data. The proposed system addresses reproducibility, interpretability, and domain adaptation gaps, offering a deployable solution for multilingual financial sentiment monitoring.
Embargo
Generative AI Based Chatbot for Customer IT Support Automation
(Institute of Electrical and Electronics Engineers Inc., 2025-12-09) Lakshan S.N; Gamage U.R; Fernando W.S.N; Kuruppu K.A.V.U.; Kodagoda, N; Abeywardhana, L
Enterprises are facing increasing challenges in delivering efficient, adaptive, and secure IT support, as traditional ticketing systems suffer from latency, rigidity, and limited scalability. Existing chatbot solutions built on commercial large language models (LLMs) often introduce high operational costs, token inefficiency, and data privacy concerns, while current retrieval and multi-agent systems lack personalization, multimodal understanding, and dynamic adaptability. This research presents a generative AI-based IT support assistant that integrates four core innovations: a multi-agent architecture enabling distributed and scalable task execution; a retrieval-augmented knowledge base enhanced with knowledge graphs and episodic memory for more accurate and personalized responses; conversational form automation using Rasa and transformer-based models to extract structured ticket information from natural queries; and an interactive 3D avatar interface with real-time speech-to-text, text-to-speech, and emotion-aware multimodal interaction to support more human-like communication. The system was evaluated using RAGAS metrics to measure accuracy, contextual recall, and response faithfulness. Results show improved domain-specific performance, reduced reliance on manual ticket handling, and lower computational overhead through caching and adaptive slot filling. This study proposes a privacy-preserving, scalable enterprise IT support framework, with future work targeting multilingual support and deeper ITSM integration.
Embargo
Multimodal Knowledge Graph for Domain-Specific Intelligence
(Institute of Electrical and Electronics Engineers Inc., 2025-06-25) Mohan, K; Munasinghe, M; Bandara, L; Wijesinghe, H; Rathnayake, S; Abeywardhana, L
In the era of information abundance, transforming vast amounts of data into meaningful knowledge remains a critical challenge, especially in domains like medicine, engineering, and education, where visual and multimodal elements play a vital role. Traditional Knowledge Graphs (KGs) excel in organizing structured and textual data but struggle to incorporate multimodal information and implicit relationships, limiting their effectiveness. This paper explores the potential of Multimodal Knowledge Graphs (MMKGs) to address these limitations by integrating text, images, videos, and audio into a unified framework. We investigate how MMKGs enhance knowledge retrieval, comprehension, and interactive learning through advanced techniques, including Natural Language Processing and deep learning. Our findings demonstrate that MMKGs significantly improve knowledge retention and application in specialized fields, offering a foundation for more intuitive and effective domain-specific knowledge ecosystems.
Embargo
Multimodal Knowledge Graph for Domain-Specific Intelligence
(Institute of Electrical and Electronics Engineers Inc., 2025) Mohan, K; Munasinghe, M; Bandara, L; Wijesinghe, H; Rathnayake, S; Abeywardhana, L
In the era of information abundance, transforming vast amounts of data into meaningful knowledge remains a critical challenge, especially in domains like medicine, engineering, and education, where visual and multimodal elements play a vital role. Traditional Knowledge Graphs (KGs) excel in organizing structured and textual data but struggle to incorporate multimodal information and implicit relationships, limiting their effectiveness. This paper explores the potential of Multimodal Knowledge Graphs (MMKGs) to address these limitations by integrating text, images, videos, and audio into a unified framework. We investigate how MMKGs enhance knowledge retrieval, comprehension, and interactive learning through advanced techniques, including Natural Language Processing and deep learning. Our findings demonstrate that MMKGs significantly improve knowledge retention and application in specialized fields, offering a foundation for more intuitive and effective domain-specific knowledge ecosystems.
Embargo
Voice Agentic Smart Reader Application In Sinhala and Tamil For Visually Impaired People
(Institute of Electrical and Electronics Engineers Inc., 2026-05-22) Siriwardana B.J.A; Hewagama A.D; Aluwihare M.C; Sharumathan M; Weerasinghe, M; Abeywardhana, L
This paper proposes an intelligent smart reading system to improve access to documents, news articles, and printed content for visually impaired and reading-challenged users, particularly in low-resource languages such as Sinhala and Tamil. Although recent advances in computer vision, speech processing, and large language models have enabled effective document understanding, existing systems often operate independently and provide limited support for non-visual navigation, interactive information access, and expressive audio output. The proposed system integrates four main components into a unified framework. First, a voice-guided navigation module enables users to capture documents without visual assistance by providing real-time audio guidance, ensuring proper alignment and reliable image acquisition for text recognition. Second, a Tamil document understanding and question answering module processes document content and supports accurate, context-aware information retrieval using a retrieval-based approach, while also providing personalized content recommendations. Third, a voice-assisted Sinhala reading module allows users to navigate documents and access information using natural voice commands and semantic processing. Finally, an emotional text-to-speech module generates expressive speech in Sinhala and Tamil, improving the naturalness and clarity of audio output. Experimental results demonstrate that the proposed system improves document capture accuracy, navigation efficiency, information retrieval performance, and speech quality. Overall, the system provides a practical, accessible, and user-centered solution for intelligent document interaction in underrepresented languages.