Verbal Question and Answer System for Early Childhood Using Dense Neural Network Method

Abstract

Questions are a well-known topic in Natural Language Processing (NLP). This feature is very suitable for use in learning activities in kindergarten to help train social interaction. The problem in this research is that the developed system must be able to understand questions from childhood. This is complex, given that their questions often need to be spoken correctly due to their limited ability to formulate questions appropriately. Therefore, this research proposes the Dense Neural Network (DNN) method, which can handle questions with non-linear word order using an Indonesian corpus of 5000 questions and answers. Experimental results show that the proposed DNN approach is superior to the Long Short Term Memory (LSTM) method in understanding and answering questions from young children, especially those that need to be more structured and formulated but have a clear context. DNN also achieved the highest accuracy in the training process, which was 0.9356. In contrast, the LSTM method showed a lower accuracy of only 0.8824. In a test of 2000 questions with different question patterns, the best accuracy was obtained by the DNN method at 93.1\%. The results of this study make an essential contribution to the development of NLP systems that can be used in the context of early childhood learning.