Abstract:
As a core task of natural language processing, text classification is widely used in many fields. Reading newspapers in daily life is a very common practice, but before reading, everyone first check the newspaper headlines. The news headline is important as it is supposed to provide an efficient way to grasp the flavor of the article and acts as a key factor to determine readers’ attitudes toward the article. In natural life, people naturally classify news in terms of themes and emotions using only headline impressions. However, in the age of constant and unorganized inflowing of digital news, manually clustering news by category and sentiment is a cumbersome job, particularly for Bangla news in which scarcity of automatic tools is found. Filling up this void, in this paper, we propose a state-of-the-art deep learning-based than we developed A Dual headed Classification model which Bangla News Headline Classification and Sentiment Analysis using transformer models. An end-to-end classification model was proposed using the BanglaBERT model to categorise (e.g., politics, religion, sports, other), and predict the sentiment polarity (positive, negative or neutral) of Bangla news headlines. Results were based on a dataset of 5033 training samples and 516 testing samples. Experimental results showed superior performance (with training accuracy of 98.18% and testing accuracy of 84.38% for aspect classification, and training accuracy of 97.17% and testing accuracy of 73.26% for sentiment analysis). Though there was a little bit of over fitting since the data set was small, the results clearly show strong potential in using pre-trained Bangla specific transformers in automated headline classification task. In the future, we will work on enlarging the dataset, further integrating data augmentation, and issuing real-time web applications for practical use of the system.
