Scholarlink

<p>The proliferation of social media has made automated classification of online discourse, such as hate speech detection and polarity prediction, an essential task for maintaining digital safety and constructive discussions. However, online conversations are complex, context-dependent, and often structured in branching discourse trees, making such tasks especially challenging. Existing classification models typically rely on limited context sampling strategies that overlook the broader structure of discussions and the semantic relevance to the target utterance. In this paper, we introduce the 2D Similarity-based Random Walk, a novel context sampling method that explores multiple paths within discourse graphs to capture more comprehensive contextual information. Through empirical analysis, we show that the proposed 2D-walk covers a wider range of branches in the discourse compared to the traditional single-path (1D) baseline, resulting in samples that are structurally richer and contain a larger number of utterances. We evaluate our method on two benchmark datasets: Guest, for detecting misogynistic hate speech, and Kialo, for predicting the polarity of argumentative replies. Extensive experiments using GPT-4, a Multi-Head Attention model, and BERT confirm that 2D-walk consistently improves classification performance across both datasets compared to several baselines, including the 1D-walk, the no-context (0D) setting, and the random sampling strategy. Our findings underscore the importance of discourse-aware sampling and suggest that leveraging both structural and semantic relationships in conversation graphs is key to advancing robust language understanding in social media contexts.</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

Modeling discourse structure with 2D similarity-based random walks for improved understanding of online conversations

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月12日

查看全文
  • Zaid Almahmoud,
  • Vibhor Agarwal,
  • Rana Mahmoud,
  • Nishanth Sastry

摘要

The proliferation of social media has made automated classification of online discourse, such as hate speech detection and polarity prediction, an essential task for maintaining digital safety and constructive discussions. However, online conversations are complex, context-dependent, and often structured in branching discourse trees, making such tasks especially challenging. Existing classification models typically rely on limited context sampling strategies that overlook the broader structure of discussions and the semantic relevance to the target utterance. In this paper, we introduce the 2D Similarity-based Random Walk, a novel context sampling method that explores multiple paths within discourse graphs to capture more comprehensive contextual information. Through empirical analysis, we show that the proposed 2D-walk covers a wider range of branches in the discourse compared to the traditional single-path (1D) baseline, resulting in samples that are structurally richer and contain a larger number of utterances. We evaluate our method on two benchmark datasets: Guest, for detecting misogynistic hate speech, and Kialo, for predicting the polarity of argumentative replies. Extensive experiments using GPT-4, a Multi-Head Attention model, and BERT confirm that 2D-walk consistently improves classification performance across both datasets compared to several baselines, including the 1D-walk, the no-context (0D) setting, and the random sampling strategy. Our findings underscore the importance of discourse-aware sampling and suggest that leveraging both structural and semantic relationships in conversation graphs is key to advancing robust language understanding in social media contexts.