You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
6 lines
1.9 KiB
6 lines
1.9 KiB
\section*{Abstract}\label{sec:abstract}
|
|
|
|
Dialog State Tracking (DST) is an essential component of task-oriented dialogue systems, which helps the dialog system understand the user's requirements for completing specific tasks. In this thesis, prompt-based methods for DST in task-oriented dialogue systems are explored by utilizing the MultiWOZ dataset. This approach does not rely on the pre-defined set of slots and their possible values. It can also be costly to label all the slots and values to train the DST models, especially for new domains. The prompt-based approach focuses on learning the DST model efficiently under low-resource few-shot settings. To examine the impact of prompt-based methods, a baseline pre-trained language model, SOLOIST\footnote{A Single Pre-trained Language Model (GPT-2) for task-oriented dialog systems}, is fine-tuned to generate belief states as a word sequence. In the prompt-based approach, the SOLOIST model is fine-tuned on limited labeled training data to generate the slots directly from values. Further, multi-prompt methods\footnote{Multi-prompt methods: \textit{Prompt Decomposition, Prompt Ensembling, Prompt Augmentation}} are applied to investigate the potential improvement in the slot generation performance. Experimental results show prompt-based methods significantly outperformed the baseline model under low-resource settings. Analysis of outputs shows prompt-based approach has some drawbacks due to the existing belief states annotation system in the MultiWOZ dataset. One limitation is that the prompt-based methods cannot generate multiple slots for repeated value candidates, as the slots are generated by passing values to the prompt function. The data, code and steps to reproduce results are publicly available \href{https://git.pavanmandava.com/pavan/master-thesis}{here}\footnote{\href{https://git.pavanmandava.com/pavan/master-thesis}{https://git.pavanmandava.com/pavan/master-thesis}}.
|
|
|
|
|
|
\clearpage |