Scholarlink

<p>The rapid expansion of digital communication has driven online social networks (OSNs) to surpass conventional news outlets, seamlessly integrating into our daily lives. This dependence on social network data has led to the emergence of various research paths, including event detection. However, such research tasks require large scale labelled social network data, and annotating such vast amounts of data is nearly impractical and often considered unfeasible. In addition, established literature encounters significant hurdles when it comes to detect unseen or new events, even after acquiring a substantial amount of training data. This marks a new era, emphasizing the importance of leveraging minimal data and embracing broader generalizations, much like human understanding of information. To this end, we proposed <Emphasis FontCategory="NonProportional">AttendFew</Emphasis>, a model to detect events in <Emphasis FontCategory="NonProportional">X</Emphasis> (formely <Emphasis FontCategory="NonProportional">X</Emphasis>) with limited data (Few-shot learning) which detects events while mitigating data dependency, even for unseen events, aligning with the current need for less data and greater adaptability. The proposed method encodes the posts (also known as “posts”) with <Emphasis FontCategory="NonProportional">BERTweet</Emphasis> and <Emphasis FontCategory="NonProportional">Graph Attention Networks</Emphasis> to capture both contextual and structural aspects of social network data and further, stacked attention will be applied to make the model attend local and global context effectively and enhances the ability to understand complex data of social network. Furthermore, the amalgamation of the feature score and Multi-layer Perceptron (MLP) facilitates class matching, aligning the derived features with their associated classifications. This operation emphasizes pivotal dimensions within the feature space while addressing data sparsity. <Emphasis FontCategory="NonProportional">AttendFew</Emphasis> is evaluated on real-world datasets and exhibits significantly better performance than state-of-the-art (SOTA) and other baseline methods in terms of accuracy, F1-score. This study represents an attempt to utilize a few-shot learning model in addressing the challenges posed by the sparsity and dynamism of online social networks for event detection.</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

An attention-enhanced few-shot model for event detection in online social networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年4月28日

查看全文
  • Sielvie Sharma,
  • Tanvir Ahmad,
  • Niyaz Ahmad Wani,
  • Naveed Ahmad,
  • Sadique Ahmad

摘要

The rapid expansion of digital communication has driven online social networks (OSNs) to surpass conventional news outlets, seamlessly integrating into our daily lives. This dependence on social network data has led to the emergence of various research paths, including event detection. However, such research tasks require large scale labelled social network data, and annotating such vast amounts of data is nearly impractical and often considered unfeasible. In addition, established literature encounters significant hurdles when it comes to detect unseen or new events, even after acquiring a substantial amount of training data. This marks a new era, emphasizing the importance of leveraging minimal data and embracing broader generalizations, much like human understanding of information. To this end, we proposed AttendFew, a model to detect events in X (formely X) with limited data (Few-shot learning) which detects events while mitigating data dependency, even for unseen events, aligning with the current need for less data and greater adaptability. The proposed method encodes the posts (also known as “posts”) with BERTweet and Graph Attention Networks to capture both contextual and structural aspects of social network data and further, stacked attention will be applied to make the model attend local and global context effectively and enhances the ability to understand complex data of social network. Furthermore, the amalgamation of the feature score and Multi-layer Perceptron (MLP) facilitates class matching, aligning the derived features with their associated classifications. This operation emphasizes pivotal dimensions within the feature space while addressing data sparsity. AttendFew is evaluated on real-world datasets and exhibits significantly better performance than state-of-the-art (SOTA) and other baseline methods in terms of accuracy, F1-score. This study represents an attempt to utilize a few-shot learning model in addressing the challenges posed by the sparsity and dynamism of online social networks for event detection.