GroupLanguage Technology and Intelligent Interaction Design

Language Technology and Intelligent Interaction Design

The group is involved in research on language technology and interaction from a variety of perspectives, ranging from empirical investigations of human-computer natural language interaction to methods and tools for development of dialogue systems. At present focus is on two projects, BCORN and ACC.

BCORN is a generic dialog strategy with domain and task-independent conventional, information-providing and recommendation capabilities. BCORN is an ambitious project aiming at developing a model for dialogue management able to handle the dialogue in any natural language interaction for any application.

The first version of BCORN is limited to conversational recommender systems and an instans of such a system, CoreSong, has been developed in the application domain of setting up an mp3-playlist.

BCORN is inspired by the work on a subsumption architecture for building intelligent creatures , as advocated by Rodney Brooks 1991. The BCORN model is constructed using dialog behaviors that each corresponds to a natural chunk of an agent's dialog strategy. Some dialog behaviors are general (e.g. a conventional dialog behavior of greeting and farewell), and some are specific (e.g. a ticket booking dialog behavior). A generic task and domain- independent dialog agent thus needs a dialog model that includes dialog behaviors that can co-exist but at the same time have a clear order of
priority. It is also imperative that the model adjusts to the needs of different back-end resources at hand in a particular application.

Similar to the model proposed by Brooks, BCORN is constructed using state automata---called dialog behavior diagrams (DBD). The DBDs express dialog behaviors of the dialog agent that are both natural conceptually and efficient computational mechanisms. The complete dialog strategy of the agent is the result of running several DBDs in parallel in a DBD strata machine, leading to an emergent coherent and flexible agent behavior.

ACC is a project with the goal of building an Automatic Contact Center. This research project is conducted in cooperation with Icepeak. Icepeak develop and sell products for call centers to a number of Swedish customers. The ACC project aims at a system that can respond automatically to FAQs. A first step is to build a database of FAQs automatically from humans calling the contact center. The database can also be used for statistical analyses of call canter traffic. In the next phase there will be a hypothesis generator helping the agent respond to questions from customers based on previous question-answers. Finally, an automatic FAQ will be developed.

In this project we develop the non-speech parts, i.e. taking textual input. The work is centered around ACC-core, a workbench for experimenting with a variety of techniques, such as LSA and decision- tree learning, for information extraction. ACC-core also includes a variety of filters, stemmers and other linguistic means for transforming data.

Within the group, there is also work done on multilingual systems, mechanized translation and term extraction in cooperation with the company Fodina, a high-tech spinnof from previous research.

Research is carried out in clode connection with the Laboratory for Natural Language Processing at Linköping University.


Arne Jönsson
+46 70 517 19 01
arne.jonsson [at]