Work

User Intent and State Modeling in \nConversational Systems

Year: 2024
Type: dissertation
Abstract: Conversational systems have become increasingly popular for their natural interaction capabilities across diverse domains. However, significant challenges persist in accurately understanding user needs and requests within multi-turn interactions. These challenges stem from the complexities of human language comprehension--an AI-complete problem, exacerbated by nuances like coreference and omission in conversational contexts. Addressing these challenges requires effective modeling of user intents and states. Towards this end, this thesis explores three fundamental tasks: user state tracking, user question rewriting, and user satisfaction estimation. \n \nEach task corresponds to one part of the thesis. The first part studies user state tracking, which involves understanding users' intentions throughout a conversation. Typically, the user state is expressed as a set of (slot, value) pairs with slots predefined and their values extracted from the conversational context. At first, this part presents a novel model that improves user state tracking by incorporating slot correlation learning with a slot self-attention mechanism. Further, this part introduces MultiWOZ 2.4, a carefully refined dataset to improve model evaluation, and two robust frameworks to improve model training by mitigating the influence of noisy state annotations. The second part delves into user question rewriting, which aims to reformulate the current user question into a self-contained and comprehensive form that preserves relevant conversational context. This part starts with introducing an advanced method that enhances user question rewriting by increasing the informativeness of question rewrites, i.e., maintaining as much relevant information from the conversational context as possible. Then, this part presents another method that enhances the quality of question rewrites by strengthening their helpfulness rather than informativeness, i.e., keeping only information helpful for downstream applications. The third part focuses on user satisfaction estimation, whose target is to predict users' satisfaction levels at each conversational turn. Estimating user satisfaction is crucial for evaluating system performance and user engagement in real time. This part proposes a pioneering method that, for the first time, considers the evolving dynamics of user satisfaction across successive turns, leading to significantly enhanced accuracy of satisfaction estimation. \n \nBy systematically advancing these three core tasks, this thesis seeks to overcome current limitations regarding user intent and state modeling and maximize the potential, controllability, and reliability of intelligent conversational systems in practical applications.
Author: Ye, Fanghua
Language: English
Open Access status: green
OpenAlex ID: W7015884180
Cites: 347
Cited by: 2
Related to: 10