Scholarlink

The purpose of this study is to predict three emotional states such as Valence, Arousal, and Dominance, based on the use of powerful language models built on transformer architecture, which have demonstrated impressive performance in several machine learning tasks such as Text Emotion Recognition (TER). To evaluate the effectiveness of these models in predicting emotional state from text, we will explore various pre-trained models such as BERT, DistilBERT, AlBERTv2, RoBERTa, and GPT2. These models are developed and applied using the IEMOCAP database. To determine the most suitable metric for evaluating Valence, Arousal, and Dominance from text, extensive experiments were conducted to compare model performance across MAE, MSE and CCC. The CCC score is the most efficient according to the results obtained in our work. Our finding advances the domain of Text emotion recognition by demonstrating the superior performance of BERT over other pre-trained transformer models in predicting continuous emotional dimensions using linguistic features and textual data.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Assessing Pre-trained Transformer Models for Textual Emotion Sensing in Continuous Space

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Awatef Messaoudi,
  • Hayet Boughrara,
  • Zied Lachiri

摘要

The purpose of this study is to predict three emotional states such as Valence, Arousal, and Dominance, based on the use of powerful language models built on transformer architecture, which have demonstrated impressive performance in several machine learning tasks such as Text Emotion Recognition (TER). To evaluate the effectiveness of these models in predicting emotional state from text, we will explore various pre-trained models such as BERT, DistilBERT, AlBERTv2, RoBERTa, and GPT2. These models are developed and applied using the IEMOCAP database. To determine the most suitable metric for evaluating Valence, Arousal, and Dominance from text, extensive experiments were conducted to compare model performance across MAE, MSE and CCC. The CCC score is the most efficient according to the results obtained in our work. Our finding advances the domain of Text emotion recognition by demonstrating the superior performance of BERT over other pre-trained transformer models in predicting continuous emotional dimensions using linguistic features and textual data.