Semantic relations classification in Hindi compound nouns using embeddings

Abstract

The present work outlines the result of Hindi compound noun interpretation using embedding features. The BERT embedding and word2vec embeddings are used to represent the features of compound nouns. The work focuses on classifying the semantic relations between the constituents of Hindi Compound nouns. The Hindi compound noun dataset used for this experiment consists of 1500 compound nouns annotated with four semantic relations. BERT-based classification is based on the model architecture of the BBC Hindi news classification task by BERT. SVM classifier with the word2vec feature is also used for another experiment. We have classified the dataset into four major semantic relations: Purpose, Modifier, Topic, and Others, based on the frequency of the relations in the dataset and then performed a multiclass classification using a BERT classifier. We achieved a significant result with 48% accuracy. We got 47% accuracy in terms of F1 score by using the word2vec model. This is one of the first experiments toward semantic relation classification of Hindi compound nouns using BERT. © 2023, The Author(s), under exclusive licence to Bharati Vidyapeeth's Institute of Computer Applications and Management.