Thursday, April 23, 15:30 - 18:00, Location: Show and Tell Area A
S. Varges, G. Riccardi, A. Ivanov, S. Quarteroni, P. Roberti
We will present a prototype dialogue system that goes beyond standard rule based models and computes on-line decisions of the best dialogue moves. The key concept of this work is that we bridge the gap between manually written dialog models (e.g. rule based) and adaptive computational models such as Partially Observable Markov Decision Processes (POMDP) based dialogue managers.
We will visualize our dialogue system through a live web-based dialogue tool that displays ongoing and past dialogue utterances, semantic interpretation confidences and distributions of confidences for incoming user acts, along with changing policies. We will demonstrate how our dedicated Spoken Language Understanding (SLU) module produces a number of candidate semantic parses using the semantics of a domain ontology and the output of Automatic Speech Recognition. We will show the internal representation of the dialogue manager (DM) including the reranking of the action set after each user utterance, and the updated values of state-action pairs based on the reward received at the end of each dialogue. We will then show a plot that relates the current dialogue's reward to the reward of previous dialogues. We will show the behavior of the DM depending on the initial condition (e.g. with pre-trained policies). These policies are obtained by interaction with user models based on data obtained from human-machine dialogues with the rule-based dialogue manager. Our strategy model allows us to experiment with many of the parameters covering the machine act. We will show the inclusion of language generation options into the action set of the dialogue manager, so that the system can explore the effect of, for example, different wordings and levels of implicit confirmation.
The system presents an example of a `thick' inter-module information pipeline architecture. Individual components exchange data by means of sets of hypotheses complemented by the detailed conversational context. All data is continuously stored in a database which the web-service based processing modules (such as SLU, DM and language generation) access. This architecture also allows us to access the database for immediate visualization.