Clause-Level Subjective Classification for Thai Article Using Bidirectional Long Short-Term Memory

Abstract

Sentence subjective classification is one of the crucial steps in analyzing opinions from such data as articles or online media which the volume has increased greatly. Extracted opinions from sentences can be used as information to produce or improve products. This research presented a method to create a model for classifying opinion at the clause level in Thai language articles using a Bidirectional Long Short-Term Memory (BiLSTM) deep learning model. This model is widely used to deal with sequential data. Moreover, the FastText model was used to convert words into numerical vectors. Our research experimented by creating models from texts in multi-domain and measuring the accuracy of the classification using the LST20 dataset. This dataset contains 44,423 pre-segmented clauses, including Part of Speech and Named Entity annotations, which are used as features for model learning. The evaluation of model performance used 5-fold cross-validation. We found that the BiLSTM model using 200 neurons in the Long Short-Term Memory unit with word and Part of Speech as features is the best model. It achieved precision of 62.562%, recall of 51.151%, accuracy score of 79.407%, and F1-score of 56.284%.