Most natural language generation is low-resource in one way or another. This talk highlights our recent efforts at Napier to deal with the challenge presented by a lack of data. I’ll start by discussing our recent efforts to develop the first corpus of conversational Scottish Gaelic for training chatbots, before shifting to an ongoing shared task aiming to collect comparable datasets for a variety of low-resource languages. Finally, I’ll share our ongoing efforts to improve the state of data-to-text generation when data is limited by using multitask learning with targeted neural NLG models.
Invited Speaker: David M. Howcroft (Edinburgh Napier University)