A framework for multilingual real-time spoken dialogue agents

Abstract

In this paper, we propose a framework for a spoken dialogue agent that is not dependent on any specific language; it takes some dialogues and sentences as training sets and uses them to acquire knowledge about the target language, then it uses this knowledge to generate several possible responses corresponding to the user input and finally it uses a simple score method to select the best one to show to the user. In aim to be language independent the system only uses very basics treatments and combines them to generate the output sentences. Moreover, all the learning and generation processes are realized in independent threads making the system enable to generate the outputs in real-time. Concretely, the user can input a new sentence at any time and influence the current output generation. We carry out experimentation in two grammaticality different languages and got some results proving our system is efficient to generate responses of a simple dialogue.