( 2019), the methodology for task-oriented dialog systems can be roughly seen to gradually progress from discriminative and modularized modeling to generative and end-to-end modeling over the recent years.Įarly methods for DST are commonly formulated as a classification task, where the dialog state representation maintains a distribution over all possible states for each slot Henderson et al. With the emergence of large-scale multi-domain TOD datasets Budzianowski et al. This encourages UBAR to adaptively supplement and make amends in response to the current user utterance in order to stay consistent and coherent during the entire session, and ultimately contribute to the task completion goal. We further propose to evaluate UBAR with the dialog context of generated content instead of the ground truth. Since in real conversations, a TOD system should be able to access the belief states it predicted and the system acts and responses it generated throughout the entire dialog session. UBAR is able to condition on the previous belief states and system acts in the dialog context, making the process of inference and planning easier for the current turn. Such training data formation resembles the workflow of a real-life task-oriented dialog session, which allows UBAR to learn task completion and language generation over the course of a dialog session. We fine-tune GPT-2 on the sequence of the entire dialog session consisting of user utterance, belief state, database result, system act, and system response of every dialog turn. To address the aforementioned limitations and advance towards a fully end-to-end TOD system, we propose UBAR to model task-oriented dialogs on a dialog session level. Third, the assumption of having access to the ground truth system responses is invalid in real conversations. Second, they use the ground truth responses from annotations in the dialog history, which makes the generation of a dialog turn independent of other turns in a dialog session. These intermediate information could be a helpful reference for the generation of the current turn. Specifically, these GPT-2-based TOD systems are trained and evaluated on a dialog turn level instead of the dialog session level, which has several limitations.įirst, the dialog history of these methods only consists of user utterances and system responses but leaves out the intermediate information such as belief states and system acts of the previous turns. In spite of the promising results from leveraging pre-trained language models like GPT-2 for end-to-end TOD systems, these methods do not fully explore the process of training and evaluating towards a real-life task-oriented dialog setting. They also incorporate database results into the training process. ( 2020a) further generalize this idea to an end-to-end setting where the belief states are also generated instead of using ground truth values. ( 2020) propose to train a unified language model for task-oriented dialogs with a single sequence in format of dialog history (all previous user utterances and responses), user utterance, belief state, system act, response of the current turn, and evaluate for DST and policy optimization. ( 2019) is shown to be capable of modeling the dialog pipeline in a unified way. On the other hand, the large pre-trained language model GPT-2 Radford et al. Visualization and a case study to illustrate the advantages of UBAR in modeling The transfer ability of UBAR to new domains with limited data and provide Thorough analyses demonstrate that the session-level training sequenceįormulation and the generated dialog context are essential for UBAR to operateĪs a fully end-to-end task-oriented dialog system in real life. Optimization, and end-to-end modeling by 4.7, 3.5, and 9.4 points respectively. Multiple settings, improving the combined score of response generation, policy The MultiWOZ datasets show that UBAR achieves state-of-the-art performances in Additionally, UBAR is evaluated in a more realistic setting, where itsĭialog context has access to user utterances and all content it generated suchĪs belief states, system acts, and system responses. Sequence of the entire dialog session which is composed of user utterance,īelief state, database result, system act, and system response of every dialog Specifically, UBAR is acquiredīy fine-tuning the large pre-trained unidirectional language model GPT-2 on the Task-oriented dialogs on a dialog session level. This paper presents our task-oriented dialog system UBAR which models
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |