Abstract
<title>Abstract</title> <p>The Community Question Answering (CQA) plat forms such as QUORA, Reddit and Stack Overflow have become important knowledge sharing sites, where users asks questions and are answered by a community of users worldwide. With the expansion of platforms, the frequency of identical and semanti cally similar questions increases tremendously. Similar questions can vary in words, structure or lexical form, although their meaning is usually the same. Such redundancy results in poor information access, moderation labor, and poor user experience. Duplicate question detection has enabled automatic processing has become a valuable research problem in Natural Language Processing (NLP) and Artificial Intelligence (AI), especially in large and dynamic systems such as QUORA. This research paper presents a hybrid deep learning framework, using Siamese Neural Networks (SNN) to detect duplicate questions. The proposed architecture comprises two parallel operating SNN sub-models; SNNA and SNNB. SNNA leverages BERT embeddings to learn semantic relationships and paraphrasing patterns, and the other sub-model SNNB captures the lexical semantics and sequential dependencies by incorporating the pre-trained Glove Embed dings, 1D-CNN and Bidirectional Long-short Term Memory (BiLSTM). Moreover, linguistic similarity characteristics are also handcrafted: word overlap, longest common subsequence (LCS), and fuzzy similarity scores are included. A feature fusion system combines both contextual, lexical and handcrafted features into one single representation, which is subsequently sent to a fully connected layer where the duplicate questions are again classified. The experiment has been carried out on the Quora Question Pairs dataset of more than 400,000 labeled question pairs, with the 80:20 stratified split forming the training and validation sets. Accuracy, precision, recall, F1-score, and confusion matrix analysis were used to evaluate model performance using AdamW optimization. The proposed architecture achieves 85.03% accuracy and 0.81 F1-Score, which is better than the baseline models. Results demonstrate that hybrid systems containing contextual and handcrafted elements enhance semantical similarity learning and facilitate duplication detection scalably.</p>