Revolutionizing Data Annotation with Convergence of Deep Learning and Active Learning to Enhance Credibility on Twitter Datasets

Active learning (AL) aims to enhance a model's performance by efficiently selecting samples for annotation, thus requiring fewer annotations overall. Conversely, deep learning (DL) can handle vast amounts of data, making them suitable for annotating large datasets quickly and efficiently enabling the model to learn and extract high-quality features. One of the most substantial obstructions in machine learning today is the integration of human-annotated data to train supervised algorithms effectively. Gathering extensive and precisely annotated datasets demands significant resources, particularly in fields such as natural language processing, encompassing tasks like speech recognition, medical imaging, knowledge extraction, text classification, and many more. Active learning addresses this issue by strategically selecting and annotating the most informative data from the unlabeled dataset. Henceforth, AL and DL highlight their pivotal role in providing effective solutions to data annotation challenge for large datasets. We have proposed a deep learning-driven active learning model that utilizes recurrent neural networks (RNNs) to annotate tweets based on credibility. We have identified and integrated sentiment polarity as the informativeness criterion, facilitating the comprehensive automation of the active learning process. The model demonstrated promising performance, as the RNN achieved an accuracy of 90%, surpassing the 78% accuracy achieved by logistic regression.