Dr. Pierre Andrews

Research in Natural Language Interfaces

Other Projects And Intersting Things

Here is a list of things I have done in the past or am doing on the side of my main research.

Semantifying Delicious

The tags2con dataset has been manually created by a group of human annotators that linked del.icio.us tags to their real meaning.

A subset of a del.icio.us dump has been used to create the tags2con dataset, a set of 500 user-bookmarks pairs have been selected randomly and all the tags used by these pairs have been manually cleaned and disambiguated to WordNet synsets.

Each tag has been split into lemmatized tokens and each of them has then been linked to its meaning in the WordNet 3.0 ontology.

We provide an RDF linked dataset of the results of this annotation effort available as dereferenceable resources.

This work has been partly supported by the INSEMTIVES project (FP7-231181).

Persuasive Computer Dialogue Improving Human-Computer Communication (PDF)

This is my Ph.D. thesis that I finished at the University of York in 2008 on Human Computer Dialogue and Persuasion.

The research reported in this thesis aims at developing techniques for achieving
persuasive dialogue between a human user and the computer; in particular,
it focuses on the formalisation of argumentation for dialogue planning in combination
with a novel framework for dialogue management to improve both the
reactivity of the dialogue system and its persuasiveness.

The main hypothesis of this thesis is that a persuasive dialogue requires specific techniques that cannot be implemented in a purely task-oriented system. In particular, Persuasive Communication requires to build social bonds, trust and involvement through social cues, as well as tailored reactions to the user’s arguments which are difficult to plan a priori. The work presented goes towards finding novel techniques for dialogue management that offer the reactivity and flexibility needed for persuasion.

In the designed dialogue framework, a layered management system that takes advantage of state-of-the-art dialogue management approaches is developed. A planning component tailored to reason about persuasion searches the ideal path in an argumentation model to persuade the user. To give a reactive and natural feel to the dialogue, a reactive component extends this task-oriented layer, using online activation techniques to select dialogue strategies that are difficult to plan a priori.

The use of planning guarantees the achievement of persuasive goals and the consistency of the dialoguewhereas the reactive component is able to better adapt to the user reactions and show more flexibility, allowing a better perception of the system by the user.

An evaluation protocol is proposed, providing a simple and independent metric for the persuasiveness of the dialogue system. The evaluation shows that the novel layered dialogue management framework – the EDEN Framework – achieves a measured persuasiveness better than a purely task-oriented system as the novel approach is able to react more smoothly to the user’s interaction. In the final evaluation, the Personage (Mairesse & Walker 2007) generator is used to display different personalities to the user during the persuasive dialogue. Experiments with the latter approach show the impact of personality simulation on the user’s perception of the interaction and show the influence of the system displayed personality on the overall persuasiveness.

Java Planning Library

This is an implementation of the GraphPlan AI planner that can be called from Java. It is implemented in Java and Prolog with all the planning code being implemented in SWI-Prolog.

The domain of the planning can be fully described in Prolog clauses and thus the planning can work on generic rules that will be extended by the theorem prover at the time of the planning.

The planner can however be easily called from Java through a clean API hiding the Prolog complexity.

The prolog code is already available on github, if you need the java libraries, please contact me.