Applications of NLG in practical conversational AI settings
Sander Wubben · 2018
Conversational AI has seen a surge in popularity recently with the rise of chatbots and smart speakers such as Amazon Alexa and Google Home. A logical extension of this development that can be expected is a rise in popularity of NLG in such settings, as the task of such a conversational system consists mainly of producing an output in natural language for a given language input. However, the extent to which NLG has seen adoption in conversational settings has been limited so far. In this talk I will give an overview to which extent NLG is used for conversational AI in the Flow.ai conversational AI platform. I will discuss training an end-to-end recurrent neural network on various data sets to build conversational systems, give an overview of the evaluation and discuss what can be expected of such systems in a practical setting. I will also discuss other uses of NLG within conversational AI, ranging from basic slot filling to more advanced functionality such as generation of data for the training of conversational AI agents or to support humans in customer service settings by providing template responses in order to decrease response times.