The paper presents a two-stage approach for emotion cause extraction in conversations, using the Flan-T5 language model.
In the first stage (THORSTATE), the model is fine-tuned to infer the emotion state of a given utterance in the conversation context. In the second stage (THORCAUSE), the model is fine-tuned to infer the emotion caused by a source utterance towards the target utterance.
The authors also propose a Reasoning Revision (RR) technique, where the model is further fine-tuned by incorporating the predicted emotion state of the source utterance to better align the state-cause dependency.
The authors analyze the training data and report quantitative statistics on the emotion-cause pairs, their distance distribution, and the correlation between speaker states and caused emotions. They also provide an algorithm-based approach for correcting the predicted emotion-cause spans.
The final submission, based on the Flan-T5base model and the rule-based span correction technique, achieves 3rd, 4th, and 5th place in the official SemEval-2024 Task 3 evaluation.
To Another Language
from source content
arxiv.org
Viktige innsikter hentet fra
by Nicolay Rusn... klokken arxiv.org 04-05-2024
https://arxiv.org/pdf/2404.03361.pdfDypere Spørsmål