Efficient Management And Optimization Of Very Large Machine Learning Dataset For Question Answering

Question answering strategies lean almost exclusively on deep neural network computations nowadays. Managing a large set of input data (questions, answers, full documents, metadata) in several forms suitable as the first layer of a selected network architecture can be a non-trivial task. In this paper, we present the details and evaluation of preparing a rich dataset of more than 13 thousand question-answer pairs with more than 6,500 full documents. We show, how a Python-optimized database in a network environment was utilized to offer fast responses based on the 26 GiB database of input data. A global hyperparameter optimization process with controlled running of thousands of evaluation experiments to reach a near-optimum setup of the learning process is also explicated.