Comparative Study of Topology and Feature Variants for Non-Task-Oriented Chatbot using Sequence to Sequence Learning

On language generation system such as chatbot and machine translation, there is a recent approach called sequence to sequence learning. This approach takes advantages of two recurrent neural networks (encoder and decoder) as an end-to-end mapping tool to generatively build the output from a certain input. In this paper, we try to find a combination of topology and feature which produces the highest result according to automatic evaluation metrics BLEU for non-task-oriented chatbot as the case study. The topologies used in the experiment are RNN, GRU, and LSTM along with their modifications, which are bidirectional encoder and attention-based decoder. The features used in the experiment are word-based feature and character-based feature. The experiment is conducted using Papaya English dialogue dataset. From the dataset, ten thousand pairs of conversation are picked for training data and a thousand pairs of conversation are picked for testing data. The result shows that bidirectional LSTM encoder with attention-based decoder and word based feature produced the highest cumulative BLEU-4 score amongst other topologies, which is 0.31.