Paraphrase detection, an essential aspect of natural language processing, involves assessing whether two provided sentences express equivalent meanings. Recent studies addressing the classification of sentence pairs for paraphrase detection commonly employ an approach that involves concatenating the embeddings of the sentences, which are then utilized as input for the classifier. This work enhances the effectiveness of paraphrase detection models by systematically and logically preparing the inputs to capture relevant features better. The effectiveness of the proposed method is confirmed through validation using two publicly available datasets, MRPC and QQP. The experimental results demonstrate superior quantitative measures, including precision, recall, and F1-measure, for the proposed approaches compared to baseline methods.