Abstract:
Lack of rich datasets is a major concern in Artificial Intelligence (AI) and Machine learning (ML). Moreover, curating proper datasets is extremely time consuming and labour intensive. Data augmentation is a solid solution to this problem. However, finding optimum augmentation magnitudes is challenging. The author has explored a model-free reinforcement learning technique called QLearning to select the best magnitudes to a given set of text data augmentation techniques. Author has taken a novel approach by utilising a model-free architecture which is easy to scale compared to existing approaches such as neural networks. With this less computationally complex solution, the author was able to provide a high accuracy rate in less time. Tests were performed on various widely used as well as custom made datasets, and it was noted that the accuracy rose from 76.8% to 97.7% when the News Analysis dataset was augmented with the proposed solution. Also it needs to be highlighted that the data points were increased by 5 times that of the original number of data points. This shows that the solution proposed by the author has been successful in achieving better accuracies and more diverse datasets, thus achieving the ambition of the project.
