Using of Recurrent Neural Network with a Refine Model to Compose Vietnamese Luc-Bat Poems for Human-Robot Verbal Interactions

Abstract

Interactions between humans and service robots require not only the machines’ friendly appearance but also the way of communication to be suitable for different cultures. Moreover, service robots that transfer information to their human partner verbally, especially in lyrical ways such as songs or poems help the conversation to be closer, more impressive, and easier to follow. This work presented an implementation containing a recurrent neural network-based model, which was trained directly on semantic verses by collecting 43,031 paragraphs for the dataset to increase the contextual creativity itself to help a service robot compose a Vietnamese traditional six-eight poetry. By also being implemented with a refining algorithm, the program can automatically replace syllables that do not follow the template rules of Vietnamese Luc bat poetry. The making of a poem is a sequential process from word to sentence, and the content of the previously complete sentence will serve as the basis for developing words for the next. Through experiments, it is found that the program performs its task impressively as the uniqueness in arranging the words goes with the diversity in the content of every generated poem.