Dokument: Simulating Intrinsic and Extrinsic User Behaviour in Task-Oriented Dialogues
| Titel: | Simulating Intrinsic and Extrinsic User Behaviour in Task-Oriented Dialogues | |||||||
| URL für Lesezeichen: | https://docserv.uni-duesseldorf.de/servlets/DocumentServlet?id=65325 | |||||||
| URN (NBN): | urn:nbn:de:hbz:061-20240325-111239-4 | |||||||
| Kollektion: | Dissertationen | |||||||
| Sprache: | Englisch | |||||||
| Dokumententyp: | Wissenschaftliche Abschlussarbeiten » Dissertation | |||||||
| Medientyp: | Text | |||||||
| Autor: | Lin, Hsien-Chin [Autor] | |||||||
| Dateien: |
| |||||||
| Beitragende: | Prof. Dr. Gasic, Milica [Gutachter] Prof. Dr. Kallmeyer, Laura [Gutachter] | |||||||
| Dewey Dezimal-Klassifikation: | 000 Informatik, Informationswissenschaft, allgemeine Werke » 004 Datenverarbeitung; Informatik | |||||||
| Beschreibung: | Task-oriented dialogue systems guide users to accomplish their goals with respect to specific tasks, such as searching for restaurants or booking flight tickets. These systems need to deal with complex and ambiguous user queries, which cannot be simplified into a sequence of keywords but are conveyed in natural language and refined through conversation. Building such systems requires substantial amounts of example dialogues for training and evaluation, which means learning and testing from real users is time-consuming and labour-intensive. As a result, user simulation is essential in developing task-oriented dialogue systems since it facilitates efficient and cost-effective interactions with such systems.
User simulators can be categorised into two groups: rule-based and data-driven. Rule-based user simulators are composed of various manually designed rules, while data-driven user simulators learn user behaviour from data. Rule-based user simulators are highly interpretable, but their use for complex scenarios is not practicable since the designing of rules may rapidly become intractable. Data-driven user simulators circumvent the need for managing complex rules in realistic scenarios but have limited explainability. There are more challenges in user simulation. Firstly, domain adaptation in user simulation is not trivial. Adapting rule-based user simulators to new domains requires rewriting handcrafted rules, which is laborious and costly. Moving data-driven user simulators to unseen domains requires retraining or even re-engineering models since they may rely on domain-dependent feature representations. Secondly, existing user models cannot capture the richness of user behaviour properly, including the user's intrinsic status, e.g. user persona and emotions, and extrinsic behaviour, e.g. user actions and utterances. These shortcomings will cause sub-optimal performance in user simulation. For instance, without considering user emotions, user models cannot capture diverse human behaviour that is driven by emotions. Generating user responses in the form of semantic actions will lose the linguistic variation in natural language, while only generating utterances in natural language may constrain interpretability. The main contributions of this thesis are as follows. Firstly, our novel proposed user simulators are domain-agnostic, overcoming the domain-dependency issue in existing user simulators and enabling transfer to unseen domains in a zero-shot fashion. Secondly, we introduce a framework for joint optimisation of user policy, which decides what actions the user model should take, and natural language generation, resulting in generating more human-like user utterances based on semantic actions and dialogue context. Thirdly, we integrate user emotions and personas into user simulation, thereby providing more fine-grained feedback beyond only task success and probing the impact of system behaviour on the user's emotional state. Our experimental results demonstrate that dialogue systems built with the help of the proposed user simulators improve user experience. In addition, modelling user behaviour in a wider spectrum, e.g. from external actions to internal status, not only allows us to explore the effects of system behaviour on the user experience but also provides a valuable tool for future research addressing ethical considerations. | |||||||
| Lizenz: |  Dieses Werk ist lizenziert unter einer Creative Commons Namensnennung 4.0 International Lizenz | |||||||
| Fachbereich / Einrichtung: | Mathematisch- Naturwissenschaftliche Fakultät » WE Informatik | |||||||
| Dokument erstellt am: | 25.03.2024 | |||||||
| Dateien geändert am: | 25.03.2024 | |||||||
| Promotionsantrag am: | 21.03.2023 | |||||||
| Datum der Promotion: | 21.02.2024 |
