Leveraging Weak Supervision and BiGRU Neural Networks for Sentiment Analysis on Label-Free News Headlines

Abstract

Auto-labeling of text is a useful and necessary technique for creating large and high-quality training data sets for machine learning models. Label-free sentiment classification is a challenging semi-supervised task in the natural language processing domain. This study leveraged the weak supervision framework to generate weak labels in three categories for millions of news headlines from Australian Broadcasting Corporation (ABC). A Bidirectional Gate Recurrent Unit (BiGRU) was then trained with neural network dense layers to achieve a validation accuracy of 96.76% with 99.99% accuracy. The performance of this method was also compared with traditional and deep learning natural language processing techniques.