14 September 2018- Ondrej Dusek: Seminar
Challenges in Response Generation and Conversational AI
This talk will focus on two recent shared task challenges I have been involved in at Heriot-Watt, both of which are concerned with response generation and conversational AI:
(1) The E2E NLG Challenge – a shared task we organized on natural language generation in task-oriented dialogue systems,
(2) The Alexa Prize – a challenge on open-domain conversational AI run by Amazon, where our team took 3rd place in 2017 and currently participates in the 2018 semi-final round.
The E2E NLG Challenge was aimed at assessing the abilities of new end-to-end approaches to natural language generation (including, but not limited to sequence-to-sequence models). We tested whether NLG systems can produce better-quality output by learning from bigger and more varied data. I will describe the challenge, summarize the results based on different system architectures, and formulate a few high-level “lessons learnt” for developing future data-driven NLG systems.
The Alexa Prize Challenge is focused on building chatbots capable of engaging and interesting social conversations on any topic. I will briefly explain the architecture of Alana, our Alexa Prize chatbot, and I will discuss several problems we faced: (1) how to avoid inappropriate replies, (2) how to select the best reply out of multiple possibilities, (3) how to deal with abuse from users, (4) how to avoid generating dull responses.
Ondrej is a Research Fellow in the Interaction Lab at Heriot-Watt University in Edinburgh, focusing on response generation in spoken dialogue systems. Before that, he got his Ph.D. from Charles University in Prague. His recent research is concerned with using deep learning methods for response generation and output quality estimation in dialogue systems. He also co-supervises Heriot-Watt University's highly successful team of Ph.D. students in the Amazon Alexa Prize chatbot competition, which reached 3rd place in 2017 and is among 8 sponsored teams out of nearly 200 applications for 2018.