Grazioso, Marco (2024) Building task-oriented dialogue systems for industry: a graph-based framework. [Tesi di dottorato]

	Testo Grazioso_Marco_36_COMPLETO.pdf Visibile a [TBR] Amministratori dell'archivio Download (7MB)
Anteprima	Testo Grazioso_Marco_36_PARZIALE.pdf Download (4MB) | Anteprima

Abstract

This thesis present the entire architecture of a task-oriented dialogue system, developed within the context of an industrial Ph.D. programme. The exploration encompasses diverse facets of the system, spanning from the training of Natural Language Understanding (NLU) models tailored for intent and entity recognition to the implementation of a dialogue management module guided by a graph representation and supported by a probabilistic system. Additionally, the dialogue state is modelled in the form of a graph, supporting different dialogue state tracking strategies. The NLU module, a fundamental component of this architecture, is composed of a hierarchical intent recognition model and an entity recognition model. This module provides its interpretations to the dialogue management module, a crucial intermediary in the system's functionality. Leveraging the configuration offered by the graph representation, the dialogue management module acquires information regarding the required slots for successful task completion. Simultaneously, the probabilistic system is strategically employed to model conditions of uncertainty, thereby activating disambiguation mechanisms. Within the conversational trajectory, the dialogue state is stored through a dedicated graph structure. This strategic choice not only facilitates efficient backward search policies for dynamically populating the necessary slots during the ongoing dialogue but also enables the linking of recognised entities to entities from external sources, such as WikiData. The representation of this information in the form of a graph enhances the system's capacity to manage and organise complex relationships within the dialogue context. In the conclusive segment of this thesis, particular emphasis will be accorded to the transition from plain text to situational dialogue. This transition is characterised by the necessity of a grounded interpretation of the user utterances. In this context, the results achieved on a benchmark specifically designed for embodied virtual agents will be shown and analysed.

Downloads

Actions (login required)

Modifica documento