Enhancing Question Answering Systems Through Optimal Hyperparameter Tuning in GRU

Deep learning is of great importance for studies in the field of natural language processing. Question Answering (QA) systems, which are widely used today, are one of these studies. QA studies are concerned with the ability to read a text and give correct answers. Many deep learning models have been tested in the field of QA. Recent studies show that Recurrent Neural Networks (RNN) are successful in this field. Proper determination and fine-tuning of hyperparameters are important in the performance of these models. In this study, hyperparameters were examined with Gated Recurrent Units (GRU), an RNN model in the QA problem. Using Stanford Question Answering (SQuAD) and NewsQA datasets, the embedding size, epoch size, batch size, vocab size, units, optimizer, corpus size and learning rate parameters were applied by changing them at 2 levels. Additionally, the effectiveness of different combinations in enhancing model performance and increasing the size of the dataset was also taken into consideration. It is seen that the interactions between the embedding size and the epochs and the optimizer play a critical role in increasing the accuracy of the model. When the results are examined, consistency is seen in both datasets.