Multi-class Short Text Classification Using Ensemble of Deep Learning Classifier

Abstract

With the substantial outgrowth of e-commerce, social media and online news portals have witnessed a great wave in expressing views through short text. Most textual contents are unstructured and messy forms, which are impractical and cumbersome to organize or manipulate by human experts. Therefore, developing an automatic short text classification model concerning low-resource languages, including Bengali, is critical. Moreover, the crucial barrier to classifying short text in Bengali is the unavailability of text corpora, scarcity of linguistics tools, a limited number of words in the text, and a lack of dependencies between the words. This paper presents a short text classification model using the ensemble of four base deep learning classifiers (Neural Network (NN), Convolutional Neural Network (CNN), Bidirectional Long Short Term Memory (BiLSTM), and Bidirectional Gated Recurrent Unit (BiGRU)). Additionally, a corpus of around 0.13 million Bengali texts is developed for short text classification into six categories (e.g., international, national, sports, amusement, technology, and politics). The evaluation results on the developed corpus demonstrated that the proposed method outperformed all the baselines machine learning and deep learning models by obtaining the highest weighted f1-score of $84.4\%$.