Scholarlink

<p>This paper presents a transformer-based semi-supervised active learning framework for tweet categorization. We focus on distinguishing <i>information</i>, <i>emotion</i>, <i>neutral</i>, and <i>irrelevant</i> tweets in case of a crisis. For this work, we considered the event “Fall of Kabul". We introduce a frequency-based sampling algorithm that deals with the temporal variations in tweet frequency for selecting tweets for initial manual annotation. A pre-trained BERT model is utilized for the classification of labelled data. As part of our iterative active learning process, we select representative tweets from the unlabelled pool. We manually annotate tweets with low-confidence scores and assign pseudo-labels to others. These labelled tweets are used to retrain the model over successive iterations. By using active learning, our approach makes full use of both supervised and unsupervised learning with reduced labelling costs while maintaining satisfactory model performance. We present the class-wise performance of the model and show that our model is able to achieve 0.806 accuracy and is effective in identifying the four classes successfully for the crisis with satisfactory F1-scores.</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

Attention-based semi-supervised active learning for multi-class tweet classification

  • 期刊论文

  • 发表日期：

    2025年6月24日

查看全文
  • Parul Khandelwal,
  • Preety Singh,
  • Rajbir Kaur,
  • Roshni Chakraborty

摘要

This paper presents a transformer-based semi-supervised active learning framework for tweet categorization. We focus on distinguishing information, emotion, neutral, and irrelevant tweets in case of a crisis. For this work, we considered the event “Fall of Kabul". We introduce a frequency-based sampling algorithm that deals with the temporal variations in tweet frequency for selecting tweets for initial manual annotation. A pre-trained BERT model is utilized for the classification of labelled data. As part of our iterative active learning process, we select representative tweets from the unlabelled pool. We manually annotate tweets with low-confidence scores and assign pseudo-labels to others. These labelled tweets are used to retrain the model over successive iterations. By using active learning, our approach makes full use of both supervised and unsupervised learning with reduced labelling costs while maintaining satisfactory model performance. We present the class-wise performance of the model and show that our model is able to achieve 0.806 accuracy and is effective in identifying the four classes successfully for the crisis with satisfactory F1-scores.