Scholarlink

<p>While numerous research efforts have focused on sentiment analysis—including an advancement from sentence to aspect-level—the complex challenge of Aspect-based Emotion Analysis (ABEA) has not been widely investigated. ABEA is a highly challenging task due to the enormous complexity of emotion classification in contrast to binary sentiments. Furthermore, rigorously labeled datasets are scarce. This paper addresses these gaps by generating a novel ABEA training dataset, consisting of 2621&#xa0;English Tweets, and fine-tuning a BERT-based model for the ABEA subtasks of Aspect Term Extraction (ATE) and Aspect Emotion Classification (AEC). The dataset annotation process made use of group annotation and majority voting strategies to facilitate label consistency. The resulting dataset contained aspect-level emotion labels for <i>Anger</i>, <i>Sadness</i>, <i>Happiness</i>, <i>Fear</i>, and a <i>None</i> class. Using the new ABEA training dataset, we developed <i>EmoGRACE</i> as a fine-tuned version of the ABSA model GRACE by Luo et al. (in: Findings of the association for computational linguistics: EMNLP 2020, pp 54–64. Association for Computational Linguistics, 2020) for ABEA. The results reflected a performance plateau at an F1-score of 70.1% for ATE and 46.9% for joint ATE and AEC extraction. Although model performance was limited by the small training dataset size and the increased task complexity, leading to overfitting and reduced generalization, the study establishes a first necessary benchmark for ABEA in social media research.</p>

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

EmoGRACE: aspect-based emotion analysis for social media data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月23日

查看全文
  • Christina Zorenböhmer,
  • Sebastian Schmidt,
  • David Hanny,
  • Bernd Resch

摘要

While numerous research efforts have focused on sentiment analysis—including an advancement from sentence to aspect-level—the complex challenge of Aspect-based Emotion Analysis (ABEA) has not been widely investigated. ABEA is a highly challenging task due to the enormous complexity of emotion classification in contrast to binary sentiments. Furthermore, rigorously labeled datasets are scarce. This paper addresses these gaps by generating a novel ABEA training dataset, consisting of 2621 English Tweets, and fine-tuning a BERT-based model for the ABEA subtasks of Aspect Term Extraction (ATE) and Aspect Emotion Classification (AEC). The dataset annotation process made use of group annotation and majority voting strategies to facilitate label consistency. The resulting dataset contained aspect-level emotion labels for Anger, Sadness, Happiness, Fear, and a None class. Using the new ABEA training dataset, we developed EmoGRACE as a fine-tuned version of the ABSA model GRACE by Luo et al. (in: Findings of the association for computational linguistics: EMNLP 2020, pp 54–64. Association for Computational Linguistics, 2020) for ABEA. The results reflected a performance plateau at an F1-score of 70.1% for ATE and 46.9% for joint ATE and AEC extraction. Although model performance was limited by the small training dataset size and the increased task complexity, leading to overfitting and reduced generalization, the study establishes a first necessary benchmark for ABEA in social media research.