DSCC Central Topic Seminar #5: Machine Learning Applications in Chatbot and Language Processing

The Data Science & Complexity Centre (DSCC) Central Topic Seminars are a series of seminars co-organized by the Utrecht Applied Data Science, the Utrecht Bioinformatics Center, and the Centre for Complex Systems Studies. It will consist of tutorials, excursions, software training and specialist lectures. We aim to expose the central topic "Machine Learning" for the broad community within and outside Utrecht University and bring the researchers from different backgrounds together.

In the first hour, Dr. Thomas Wolf, Science lead at Hugging Face, will give a specialist lecture on Machine Learning Applications in Chatbot titled "Recent developments in Neural Network Based Dialogue Agents". (Slides)

Abstract: Free-form dialogue systems (also called chatbots) are dialog agents that are designed to interact with humans in open-ended conversations ("small talk"). Developing these systems tackles the general research question of how a model can generate a coherent and interesting discussion with a human-being. These dialog agents are thus test-beds for many interactive AI systems and are also directly useful in applications ranging from technical support services to entertainment. However, building such intelligent conversational agents remains an unsolved problem in artificial intelligence research. I will present a summary and the technical details of our winning participation to the Conversational Intelligence Challenge 2 which was part of the NeurIPS 2018 conference held in Montreal ( in early December 2018. The Conversational Intelligence Challenge aimed at testing how an agent could be provided with a simple personality and common sense reasoning abilities to generate meaningful responses. Our agent showed strong improvements on all tested metrics topping the leaderboard with a significant margin over the second top model. These improvements were obtained by using a technique called transfer learning to incorporate common-sense knowledge learned from a large corpus combined with a multi-tasks training scheme that was able to take advantage of orthogonal inductive learning biases.

In the second hour, Joost Bastings from the Institute for Logic, Language and Computation of the University of Amsterdam will give a specialist lecture on Machine Learning Applications in Natural Language Processing titled "Linguistically informed Neural Networks for NLP". (Slides)

Abstract: In recent years many natural language processing (NLP) pipelines were replaced by deep neural networks, often with state-of-the-art performance as a result. In the process, the linguistic knowledge that we used to leverage to solve tasks (for example parse trees) was left behind and traded in for neural feature extractors that simply take words (or characters) as input, and nothing else. I will present how graph-convolutional networks (GCNs) can be used to incorporate linguistic knowledge (syntax and/or semantics) in deep neural networks for the tasks of Semantic Role Labeling and Neural Machine Translation, with modest improvements as a result.

Both students and staff are welcome.

DSCC Central Topic Seminars:

Some of the seminars are available to watch via the CCSS YouTube channel.

Venue: Room 2.01, Minnaert Building, Leuvenlaan 4, De Uithof, Utrecht

Please register before Thursday 10 January 2019.

Start date and time
End date and time
Minnaertgebouw 2.01