The Hybrid BERT-LSTM Model for the classification Sindhi Text in NLP

Abstract

The traditional ML models lack to capture the relationship having deep semantic nature, while deep learning model alone cannot work better with temporal and contextual embeddings. In this context the need of efficient Hybrid approach BERT-LSTM for the improvement of the text classification. This study proposes the Hybrid approach BERT-LSTM on the sindhi text data. The text data is collected in sindhi language from hugging face. The dataset contains the labeled samples of the sindhi language text having their predefined classes. Total 150 sentences are used for the sindhi text classification. The model performed robust performance results by the all-evaluation matrices, which achieved macro-average of 0.88, 0.88 accuracy and 0.86 precision and recall 0.85. the significant use of the macro-average because it confirms the consistent model predictive ability across the sentimental data textual classes. The BERT embeddings provide sustainable granularity in sindhi text syntax might provide the miss classification with is shown in sense of minimal dispersion off-diagonal cells. This study provides the critical gap in sentimental analysis for the sindhi text data by providing the hybrid approach BERT-LSTM model architecture. The multilingual BERT is provided to add for feature extraction and for the modeling capability for the sequential capability the Bidirectional BERT is used. The semantic nuance and the low of sindhi text structural behavior is effectively captured by the Hybrid approach.