Word Embedding based Textual Semantic Similarity Measure in Bengali

Textual semantic similarity is a crucial constituent in many NLP tasks such as information retrieval, machine translation, information retrieval and textual forgery detection. It is a complicated task for rule-based techniques to address semantic similarity measures in low-resource languages due to the complex morphological structure and scarcity of linguistic resources. This paper investigates several word embedding techniques (Word2Vec, GloVe, FastText) to estimate the semantic similarity of Bengali sentences. Due to the unavailability of the standard dataset, this work developed a Bengali dataset containing 187031 text documents with 400824 unique words. Moreover, this work considers three semantic distance measures to compute the similarity between the word vectors using Cosine similarity with no weight, term frequency weighting and Part-of-Speech weighting. The performance of the proposed approach is evaluated on the developed dataset containing 50 pairs of Bengali sentences. The evaluation result shows that FastText with continuous bag-of-words with 100 vector size achieved the highest Pearson's correlation (rho) score of 77.28%. (C) 2021 The Authors. Published by Elsevier B.V.