Can explicit task models improve dialogue understanding?

February 15, 2018
Versione stampabile

Time: 11.00 a.m.
Location: Ofek Room, Polo scientifico e tecnologico "Fabio Ferrari", Building Povo 1, via Sommarive 5, Povo (Trento)

Speaker

  • Prof. Bernd Ludwig, University of Regensburg

Abstract

In recent years, approaches to dialogue understanding primarily rely on stochastic or deep learning models trained on big corpora of domain typical dialogues.

In many situations, however, it is hard to apply such approaches for new dialogue domains, for which no corpora exist. In order to address this issue, in my talk I try do analyse a dialogue domain in a communicative and a task related part. I will model both parts with Concurrent Task Trees (CTT) that encode plans in the domain and for communication about domain related tasks.

With such an analysis, the semantic interpretation of utterances in a dialogue can be learned by (1) learning typical utterances for tasks defined in a CTT, (2) learning the communicative intention of utterances, and (3) learning the discourse function of utterances in the context of a current task.

I present a tool that we have developed to collect dialogue corpora in cooperative domains and show examples for natural language interaction with a pedestrian navigation system. Finally, I present our plans to collect corpora in the same way for dialogues of phobia patients with a virtual agent that serves as a psychotherapist and for dialogues of flight passengers with an service assistant in the Munich airport.

About the Speaker

Prof. Bernd Ludwig studied Computer Science at Friedrich-Alexander-University in Erlangen-Nuremberg from 1992-1997. His 2004 phd thesis focused on plan-based dialogue management. Bernd finished his habilitation at the same university in 2010 (title of thesis: Plan-based HCI in multimodal recommender systems ("Planbasierte Mensch-Maschine-Interaktion in multimodalen Assistenzsystemen").

1997-2004 Bernd was a post-doctoral researcher at the knowledge processing group of the Bavarian Research Center for Knowledge-Based Systems in Erlangen. He became an assistant professor at the Chair of Artificial Intelligence (Friedrich-Alexander-University in Erlangen-Nuremberg) in 2004. In 2010, Bernd was appointed BIT guest professor und postdoc researcher at the Free University of Bozen and the Department of Information Engineering and Computer Science der Universitá degli Studi di Trento. After his return in November 2010, he was an associate professor (Privatdozent)  until March 2011. Bernd was appointed as professor of Computational Linguistics at the Chair of Information Science at University of Regensburg in August 2011. His research focuses on the dependencies of Communicative Action in Human Computer Interaction as well as its implementation using efficient algorithms. Furthermore, he is interested in the use of mobile devices as an aid to solve complex tasks using recommender systems.

Contact Person for this Seminar: Prof. giuseppe.riccardi [at] unitn.it (Giuseppe Riccardi)