Year: 2023
Type: dissertation
Abstract: Emotion Recognition in Conversation (ERC) is an important task in Natural Language Processing.Although ERC has been intensively investigated, most of previous studies fail to full exploit the knowledge within language models.Furthermore, to the best of our knowledge, ERC has been never investigated under few-shot or zero-shot settings.In this thesis, we propose prompt-based finetuning approaches to full-shot and few-shot ERC.We combine intra-and inter-speaker dependencies in dialogues within these approaches.Experiment results on widely-used benchmarks show that our proposed approaches achieves better or comparable performance to SOTA baselines under the fullshot setting, and outperform these baselines by a large margin under few-shot settings.Additionally, we apply in-context learning for zero-shot ERC.We propose an effective strategy for the generation of demonstration examples, analyze behaviors of the large language model through experiments, and summarize good practices in prompt design for dialogues.
Author: Chen, Zikai
Institution: The University of Tokyo
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7146205996