Document Type

Theses, Ph.D

Disciplines

Computer Sciences

Publication Details

Thesis submitted for the degree of Doctor of Philosophy, Technological University Dublin, January 2023.

Abstract

Human-machine conversational agents have developed at a rapid pace in recent years, bolstered through the application of advanced technologies such as deep learning. Today, dialogue systems are useful in assisting users in various activities, especially task-oriented dialogue systems in specific dialogue domains. However, they continue to be limited in many ways. Arguably the biggest challenge lies in the complexity of natural language and interpersonal communication, and the lack of human context and knowledge available to these systems. This leads to the question of whether dialogue systems, and in particular task-oriented dialogue systems, can be enhanced to leverage various language properties. This work focuses on the semantic structural properties of language in task-oriented dialogue systems. These structural properties are manifest by variable dependencies in dialogue domains; and the study of and accounting for these variables and their interdependencies is the main objective of this research.

Contemporary task-oriented dialogue systems are typically developed with a multiple component architecture, where each component is responsible for a specific process in the conversational interaction. It is commonly accepted that the ability to understand user input in a conversational context, a responsibility generally assigned to the dialogue state tracking component, contributes a huge part to the overall performance of dialogue systems. The output of the dialogue state tracking component, so-called dialogue states, are a representation of the aspects of a dialogue relevant to the completion of a task up to that point, and should also capture the task structural properties of natural language. Here, in a dialogue context dialogue state variables are expressed through dialogue slots and slot values, hence the dialogue state variable dependencies are expressed as the dependencies between dialogue slots and their values. Incorporating slot dependencies in the dialogue state tracking process is herein hypothesised to enhance the accuracy of postulated dialogue states, and subsequently potentially improve the performance of task-oriented dialogue systems.

Given this overall goal and approach to the improvement of dialogue systems, the work in this dissertation can be broken down into two related contributions: (i) a study of structural properties in dialogue states; and (ii) the investigation of novel modelling approaches to capture slot dependencies in dialogue domains.

The analysis of language's structural properties was conducted with a corpus-based study to investigate whether variable dependencies, i.e., slot dependencies when using dialogue system terminology, exist in dialogue domains, and if yes, to what extent do these dependencies affect the dialogue state tracking process. A number of public dialogue corpora were chosen for analysis with a collection of statistical methods being applied to their analysis.

Deep learning architectures have been shown in various works to be an effective method to model conversations and different types of machine learning challenges. In this research, in order to account for slot dependencies, a number of deep learning-based models were experimented with for the dialogue state tracking task. In particular, a multi-task learning system was developed to study the leveraging of common features and shared knowledge in the training of dialogue state tracking subtasks such as tracking different slots, hence investigating the associations between these slots. Beyond that, a structured prediction method, based on energy-based learning, was also applied to account for explicit dialogue slot dependencies.

The study results show promising directions for solving the dialogue state tracking challenge for task-oriented dialogue systems. By accounting for slot dependencies in dialogue domains, dialogue states were produced more accurately when benchmarked against comparative modelling methods that do not take advantage of the same principle. Furthermore, the structured prediction method is applicable to various state-of-the-art modelling approaches for further study.

In the long term, the study of dialogue state slot dependencies can potentially be expanded to a wider range of conversational aspects such as personality, preferences, and modalities, as well as user intents.

DOI

https://doi.org/10.21427/XCMR-7N05

Creative Commons License

Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License
This work is licensed under a Creative Commons Attribution-NonCommercial-Share Alike 4.0 International License.

Download
Contact the Author


Share

COinS