Publication

Abstract : Dependency parsing is concerned with obtaining its grammatical structure and establishing the relations between the words in the sentence, namely the headwords and the words that alter those heads. This paper presents a transformer-based dependency parser for Hindi. The transformer model, BERT and sentence transformer LaBSE were used to generate the embeddings for the Hindi tokens, further applied to the machine learning algorithms to make predictions. The accuracy and the Unlabelled attachment score (UAS) are calculated for the models using Multilingual BERT cased and uncased (MBERT), Multilingual Representations for Indian Languages (MuRIL) and LaBSE. The performance in the case of MBERT ranges between 55 and 65%, while the results observed on using the LaBSE are 3–4% higher for UAS score and 2–3% greater for the accuracy of most models. The Decision Tree algorithm provides the UAS score of 61% with MBERT and 65% with LaBSE, and Random Forest provides a score of 63% and 66% with MBERT and LaBSE, respectively. The UAS score offered by SVM for three kernels–Linear, RBF, and Polynomial–was obtained to be 63%, 60% each, respectively, using the MBERT, while the scores 64% 66% and 66.50% were obtained using the LaBSE embedding. The Logistic Regression is observed to provide a score of 62% for MBERT and 63.7% for LaBSE.