fastText-Based Siamese Network for Hindi Semantic Textual Similarity

Abstract

Semantic textual similarity is a measurement of the degree of similarity or equivalence between two sentences semantically. Semantic sentence pairs have the ability to substitute text from each other and retain their meaning. Various rule-based and machine learning models have gained quick prominence in the field, especially in a language like English, where there is an abundance of lexical tools and resources. However, other languages like Hindi have not seen much improvement in state-of-the-art methods and models to evaluate semantic similarity of text data. This paper proposes a fastText-based Siamese neural network architecture to evaluate the semantic equivalency between a Hindi sentence pair. The pair is scored on a scale of 0–5, where 0 indicates least similar and 5 indicates most similar. The corpus contains a combination of two datasets containing manually scored sentence pairs. The performance parameters used to evaluate this approach are model accuracy and model loss over a training period of multiple epochs. The proposed architecture incorporates a fastText-based embedding layer and a bi-directional Long Short Term Memory layer to achieve a similarity score. The proposed architecture can extract semantic and various global features of the text to determine a similarity score. This model achieves an accuracy of 85.5% on a compiled Hindi-Hindi sentence pair dataset, which is a considerable improvement over existing rule and supervise-based systems. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.