Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Computing | en_US |
| dc.contributor.advisor | Li, Wenjie Maggie (COMP) | en_US |
| dc.creator | Wang, Jian | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/13394 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Towards proactive dialogue systems with target : planning to generation | en_US |
| dcterms.abstract | Building intelligent dialogue systems has been a long-standing goal in Artificial Intelligence (AI). Over recent years, dialogue systems have been primarily developed to entertain people or accomplish specific tasks. Though significant progress has been made, there is still a big gap between previous dialogue systems and people’s expectations. Existing social dialogue systems typically converse passively, struggling to actively maintain engaging conversations. Although the models developed for task-oriented dialogues, such as booking flights or hotels, may ask questions to collect or even clarify the information needed, they are mainly concerned with completing the user-side goals. However, many application scenarios in the real world, such as targeted product promotion, intelligent tutoring, and psychological consultation, require the system to achieve its own goal through proactive dialogues. In such scenarios, a qualified dialogue system is expected to proactively lead a conversation toward the goal, such as making targeted recommendations, delivering predesignated knowledge, providing emotional support, and so forth. It is essential to endow dialogue systems of this type with the capability to lead conversations with users proactively and strategically to achieve the system-side goal. | en_US |
| dcterms.abstract | In this thesis, we aim to explore target-oriented (or goal-directed) proactive dialogue systems. After being designated a target in the form of <dialogue act, topic>, a system’s objective is to lead the dialogue towards the designated target proactively and coherently. When appropriate, the system needs to achieve the target act (e.g., recommendation) on the target topic. In particular, we identify three main research questions (RQs): (1) How to plan appropriate dialogue acts and topics that lead the dialogue to reach the target proactively? (2) How can we achieve more coherent planning with smooth transitions throughout dialogue by considering the user’s feedback? (3) How to leverage planning to guide system utterance generation and how can generated utterances be consistent with the system’s goal-directed role? To address these critical yet challenging issues, we take a divide-and-conquer approach by decomposing the target-oriented proactive dialogue task into two sub-problems: dialogue planning and dialogue generation. According to the category of the work conducted, this thesis is divided into three parts. | en_US |
| dcterms.abstract | In the first part (works 1 and 2), we systematically formulate the target-oriented proactive dialogue setting and aim to address RQ-1. We define the “target” as a pair of <dialogue act, topic> designated to the system, where we require the system to proactively lead conversations to achieve the specific act on the designated topic at an appropriate time. Then, we build a basic plan-then-generate pipeline, where we plan a dialogue path comprising a sequence of <dialogue act, topic> pairs, grounding on which we generate system utterances in a prompt-based manner. Furthermore, we propose a novel target-constrained bidirectional planning (TRIP) approach inspired by decision-making theories in cognitive science. We plan an appropriate dialogue path from both forward and backward directions, employing two individual language decoders to supervise each other. They are trained to converge on consistent <dialogue act, topic> pairs by minimizing the decision gap. During the inference stage in planning, we adopt a target-constrained decoding strategy with a bidirectional agreement to control the planning. | en_US |
| dcterms.abstract | The second part (work 3) studies solutions for RQ-2 under the setting we formulated. Beyond the target, recognizing the user’s willingness or feedback to follow the system is crucial for achieving coherent transitions. This motivates us to incorporate underlying user feedback into planning as well as maintaining the system’s goal-directed behavior. Inspired by the good properties of the Brownian bridge stochastic process in coherence control, we propose a coherent dialogue planning approach called Color. The user feedback is incorporated to perturb the density and uncertainty of the Brownian bridge stochastic process, simulating its impact on the system’s target-oriented behavior. Our training utilizes a contrastive objective, enabling COLOR to retain global coherence. Extensive experiments suggest that the proposed method effectively improves user-aware coherence in the dialogue planning sub-problem. | en_US |
| dcterms.abstract | In the third part (works 4 and 5) of this thesis, we mainly delve into RQ-3 and investigate feasible and efficient methods for dialogue generation. We take the planned dialogue paths as guidance and devise a plan-controlled generation approach, which shifts the hidden states of generation in the desired direction (i.e., the planned path) with specified strength through gradients. Furthermore, we observe an emergent need for high-quality datasets, while building one from scratch requires tremendous human effort. We thereby propose an automatic dataset curation framework using a role-playing approach, then construct a large-scale target-oriented dialogue dataset named TOPDIAL to largely boost related research. Since the system’s target-oriented behavior is long-term, it is non-trivial to maintain the role consistency that adheres to its goal as the dialogue progresses. We propose an efficient Multi-round Interactive Dialogue Tuning (Midi-Tuning) framework, with two adapters built upon LLMs and a round-level memory caching mechanism. Extensive experiments show that our framework performs superior to traditional fine-tuning. | en_US |
| dcterms.abstract | In summary, this thesis comprehensively studies target-oriented proactive dialogue systems, investigating essential aspects of dialogue planning and utterance generation. We showcase the effectiveness of the proposed approaches, indicating the great potential of our work towards building intelligent dialogue systems that are applicable in the real world. | en_US |
| dcterms.extent | xviii, 205 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2024 | en_US |
| dcterms.educationalLevel | Ph.D. | en_US |
| dcterms.educationalLevel | All Doctorate | en_US |
| dcterms.LCSH | Artificial intelligence | en_US |
| dcterms.LCSH | Human-computer interaction | en_US |
| dcterms.LCSH | Question-answering systems | en_US |
| dcterms.LCSH | Hong Kong Polytechnic University -- Dissertations | en_US |
| dcterms.accessRights | open access | en_US |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13394