Deep Learning-Based Parts-of-Speech Tagging in Marathi Language
Pallavi R. Deore, Nita V. Patil, Ajay S. Patil · Procedia Computer Science · 2025
The proposed research investigates a novel approach for Parts of speech tagging in the Marathi language that uses a deep learning-based approach. This task contributes to the Parts of Speech tagging process, a preliminary stage in numerous natural language processing programs and algorithms that attribute grammatical labels to each word in Marathi phrases. Earlier methods for parts of speech (POS) tagging were based on rules and heuristics. Over time, Machine learning techniques like artificial neural networks, were introduced to improve accuracy. Recently, deep learning has made POS tagging even more accurate, especially for high-resource languages like English. However, low-resource languages like Marathi still face challenges in developing efficient and precise POS tagging methods. This work addresses that gap by proposing a deep learning-based approach for POS tagging in Marathi. The corpus comprises 48,420 annotated words. An 80:20 split of the corpus was made, with 80% going toward training the model and the remaining 20% going toward testing. The training set was then split up into training and validation sets. The objective of this investigation was to compare and implement a variety of deep learning-based components in Marathi speech taggers, such as Recurrent Neural Networks (RNN), gated recurrent units (GRU), Long Short-Term Memory (LSTM), and Bidirectional Long Short-Term Memory (BiLSTM). We experimented with 4, 16, 32, and 64 hidden states with 30, 50, and 100 epochs for each hidden state to determine the best setup for each model. The proposed approach demonstrated satisfactory performance for the RNN, LSTM, GRU, and BiLSTM models, The RNN model achieved a precision of 97.43%, a recall of 97.49%, an F1 score of 97.42%, and a test accuracy of 97.49%. The LSTM model performed better, with a precision of 97.53%, recall of 97.55%, an F1 score of 97.50%, and an accuracy of 97.55%. The GRU model yielded a precision of 97.76%, a recall of 97.76%, an F1 score of 97.69%, and an accuracy of 97.69%. Finally, the Bi-LSTM model achieved the highest performance with a precision of 97.96%, recall of 97.86%, an F1 score of 97.78%, and an accuracy of 97.86%. The results demonstrate that the BiLSTM outperforms all other models.