Enhancing Paraphrase Evaluation in Marathi Question Answering Systems Using Similarity Techniques

Abstract

In Natural Language Processing applications, paraphrasing is essential, especially in Question Answering (QA) systems where users may provide distinct yet valid responses to the same questions. This study presents a structured approach for evaluating paraphrased answers in the Marathi language using multiple similarity measures, including Levenshtein Distance, Jaccard Similarity, and Cosine Similarity. The proposed methodology integrates one-to-one word matching, masking techniques, synonym dictionary verification, and dependency parsing to ensure grammatical and syntactic consistency. A dataset comprising 540 questions from the Balbharti Standard 2 textbook, each with three student-generated paraphrased answers, was analyzed. The findings indicate that this approach effectively captures lexical and semantic similarities, enhancing the robustness and fairness of Marathi QA systems. By leveraging multiple similarity measures, this study establishes a systematic framework for paraphrase identification, improving automated answer evaluation.