http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/4862 Tue, 28 Apr 2026 07:52:18 GMT 2026-04-28T07:52:18Z http://repo:8080/xmlui/bitstream/id/31730475-a8ac-4893-8967-24d40f58055e/ http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/4862 http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/4873 Title page Wed, 19 Feb 2025 00:00:00 GMT http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/4873 2025-02-19T00:00:00Z http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/4872 Front materials Wed, 19 Feb 2025 00:00:00 GMT http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/4872 2025-02-19T00:00:00Z http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/4871 A Comparative Analysis of Deep Learning Algorithms for Formality Classification in Texts Using Linguistic Features Karunarathna, K.M.G.S.; Rupasingha, R.A.H.M.; Kumara, B.T.G.S. Because of the wide variety of formal and informal writing styles brought about by the rapid growth of digital communication, the classification of documents based on it becomes a challenging task. Using a variety of variables, this work seeks to increase the accuracy of formality classification algorithms. Grammar, vocabulary, punctuation, and sentence structure are some stylistic components that define various writing styles, and traditional approaches have trouble distinguishing between them. Differentiating between formal and informal language is becoming increasingly important in applications such as research papers, legal documents, informal letters, NEWS, etc. The objective of this approach is to use linguistic features, examines how well deep learning algorithms classify documents as formal and informal. The study collected dataset of 5,000 text samples. The text files contained 2500 formal letters, news items as formal documents, and remaining are personal blogs, personal letters as informal documents. Next pre-processed all data using stop word removal, lemmatization, tokenization and lowercasing. Formal and informal categories which include pronouns, grammar, vocabulary, slang, acronyms, language and initialisms seven linguistic features were targeted for this study and those features are extracted. Then these seven features are combined to generated the feature vector for each document. The generated feature vector was applied and in order to classify documents, three deep learning models Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks are trained. Here, ANN learns nonlinear patterns in data, CNN identifies text sections, and LSTM considers word position and those are selected based on the literature review. The performance of each model is compared using different test splitting methods and cross-validation techniques. According to experimental data, the LSTM model outperforms ANN and CNN in terms of precision, recall and f measure metrics, achieving the highest classification accuracy of 89.4% with an epoch size of 100 and a batch size of 32 with lowest error rate for Mean Absolute Error and Root Mean Squared Error. The results highlight how well LSTM can detect linguistic subtleties and offer suggestions for improving formality recognition in Natural Language Processing applications, which will help with more context sensitive text classification. Wed, 19 Feb 2025 00:00:00 GMT http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/4871 2025-02-19T00:00:00Z http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/4870 Towards Ethical Inference in Language Models: Integrating Religious Data and Enhancing Responsible LLM Development Ranasinghe, K.S.; Paik, I. Large Language Models (LLM) have emerged as the most powerful tools to perform various aspects in daily life. These models are capable of diverse tasks including text understanding and generating, image generation, language translation and sentiment analysis. Continual advancements in LLM are expanding the scope of their capabilities enabling wide range of applications. Although LLMs have made significant progress, still there are challenges and limitations that needs to be addressed. As the existing LLM models generally focus on the natural language processing related tasks, it is crucial to emphasize the training and fine-tuning of ethical LLMs. When developing and fine-tuning LLMs, issues such as biased responses and lack of moral consistency can arise. This could lead to significant ethical challenges, particularly because the data used for training heavily influences the model’s outputs. Developing a specific ethical LLM by establishing a benchmark for ethical performance could help overcome this problem. The primary goal of this research is to implement an ethical inferences language model which can make the predictions based on the religious data. Religious data is used for the fine-tuning and Llama-2-7B-chat model is used along with Low Rank Adaptation techniques. The fine-tuned model was tested by generating the responses to prompts related to ethical scenarios and the accuracy of the model can be calculated. The model trained with 5000 Bible data. During the training loss decrease gradually by denoting the model learns well with the data. The fine-tuned model provides reliable performance when working with ethics-related data. Further the Fine-tuned model demonstrated the ability to generate text based on ethical prompts, showing a positive trend in the generated ethical inferences indicating that this model can be developed further by training with more religious data from Bible, Quran, Hindu scriptures and Tripitaka. In future the model will be refined further using Supervised Fine Tuning to obtain more accurate model with enhanced ethical inference capabilities. Wed, 19 Feb 2025 00:00:00 GMT http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/4870 2025-02-19T00:00:00Z