Skip navigation

DSpace

機構典藏 DSpace 系統致力於保存各式數位資料(如:文字、圖片、PDF)並使其易於取用。

點此認識 DSpace
DSpace logo
English
中文
  • 瀏覽論文
    • 校院系所
    • 出版年
    • 作者
    • 標題
    • 關鍵字
    • 指導教授
  • 搜尋 TDR
  • 授權 Q&A
    • 我的頁面
    • 接受 E-mail 通知
    • 編輯個人資料
  1. NTU Theses and Dissertations Repository
  2. 管理學院
  3. 資訊管理學系
請用此 Handle URI 來引用此文件: http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/106052
標題: HateKG:針對仇恨言論偵測的屬性增強仇恨言論知識圖譜
HateKG: An Attribute-Enriched Hate Speech Knowledge Graph for Hate Speech Detection
作者: 林佳穎
Chia-Ying Lin
指導教授: 陳建錦
Chien-Chin Chen
關鍵字: 仇恨言論偵測; 隱性仇恨言論; 知識圖譜; 屬性增強知識圖譜; 檢索增強推理; 大型語言模型
Hate Speech Detection; Implicit Hate Speech; Knowledge Graph; Attribute-Enriched Knowledge Graph; Retrieval-Augmented Reasoning; Large Language Models
出版年 : 2026
學位: 碩士
摘要: 隨著社群媒體快速發展,仇恨言論得以廣泛傳播,並且可能對社會造成負面影響。現有的仇恨言論偵測模型通常能夠辨識包含辱罵、歧視性詞彙或具有攻擊性的顯性仇恨言論,但對於隱性仇恨言論的辨識仍具有挑戰。隱性仇恨言論通常不會使用明確的仇恨字詞,而是透過刻板印象、暗示、諷刺、隱喻或背景知識傳達具有傷害性的意圖。因此,僅依賴輸入文字進行判斷,可能無法充分理解訊息背後隱含的語意,進而產生錯誤的預測結果。

為了解決此問題,本研究提出 HateKG,應用於仇恨言論偵測的屬性增強仇恨言論知識圖譜。本研究利用大型語言模型從仇恨言論資料集中擷取三元組,以結構化的形式表示文字中與仇恨相關的隱含語意。為了提升知識圖譜的品質,本研究進一步設計三元組驗證與正規化流程,過濾與修正不適當的三元組,並將語意相近的實體與關係合併為一致的標準化表示。除了圖譜結構之外,本研究亦為實體與關係加入多種屬性,包括實體層級的情感極性、關係層級的情緒資訊,以及關係特異性。這些屬性可提供額外的語意與情感訊號,能夠更完整的表示與仇恨言論相關的語境。

在偵測階段,本研究將 HateKG 作為 IKAR(Iterative Knowledge-Augmented Reasoning)框架中的外部知識來源。給定一段輸入文字後,模型會從 HateKG 中擷取與輸入內容相關的三元組,並將其作為外部證據提供給大型語言模型進行推理。透過此方式,模型在進行預測時,不再僅依賴文字表面的資訊,而能進一步結合結構化的仇恨相關知識作為判斷依據。

本研究於 HateXplain、Latent Hatred 與 ToxiGen 三個資料集上進行實驗。實驗結果顯示,相較於直接使用 GPT-5.1 進行分類,以及經微調的 BERT 基準模型,IKAR 在三個資料集上皆取得最佳的 F1 分數。此結果顯示,結構化的外部知識能夠提升仇恨言論偵測的效果,尤其是在需要背景知識與語境推理的隱性仇恨言論情境中。此外,本研究亦針對情感極性屬性進行消融實驗。結果顯示,情感極性能夠提升 ToxiGen 與 HateXplain 上的模型表現,但在 Latent Hatred 上則造成些微下降。進一步的案例分析顯示,實體層級的情感極性可以作為弱檢索訊號,但其表示方式仍較為粗略,無法充分區分一個與仇恨相關的概念是在被認同、批判、引用,或僅被提及。

總結來說,本研究證明屬性增強知識圖譜能夠為隱性仇恨言論偵測提供有效的外部證據。透過將仇恨相關的隱含語意組織為結構化三元組,並加入語意與情感相關屬性,HateKG 能夠協助模型突破僅依賴文字表面線索的限制,進一步提升隱性仇恨言論偵測的推理能力與可解釋性。
With the rapid development of social media, hate speech can spread widely and cause harmful social impacts. Although existing hate speech detection models can often identify explicit hate speech that contains direct insults, discriminatory terms, or offensive expressions, they still struggle with implicit hate speech. Implicit hate speech usually does not rely on explicit hateful keywords. Instead, it conveys harmful intent through stereotypes, implications, sarcasm, metaphors, or background knowledge. Therefore, models that only rely on the input text may fail to capture the hidden meaning behind the message and make incorrect predictions.

To address this problem, this study proposes HateKG, an attribute-enriched hate speech knowledge graph for hate speech detection. HateKG is constructed from hate speech datasets by using a large language model to extract subject--predicate--object triplets, which represent hate-related implicit meanings in the text. To improve the quality of the knowledge graph, this study further designs triplet validation and triplet normalization procedures to filter and repair unsuitable triplets and merge semantically similar entities and relations into consistent canonical representations. In addition to the graph structure, this study enriches entities and relations with attributes, including entity-level polarity, relation-level emotion, and relation specificity. These attributes provide additional semantic and affective signals for representing hate-related contexts.

During detection, HateKG is used as the external knowledge source in IKAR (Iterative Knowledge-Augmented Reasoning). Given an input text, the framework retrieves relevant triplets from HateKG and provides them as external evidence for large language model reasoning. In this way, the model can make predictions not only based on the surface text, but also with support from structured hate-related knowledge.

Experiments are conducted on three datasets: HateXplain, Latent Hatred, and ToxiGen. The results show that IKAR achieves the best F1 score across all three datasets compared with direct GPT-5.1 classification and fine-tuned BERT-based baselines. This indicates that structured external knowledge can improve hate speech detection, especially in implicit hate speech scenarios that require background knowledge and contextual reasoning. In addition, this study conducts an ablation study on the polarity attribute. The results show that polarity improves performance on ToxiGen and HateXplain, but slightly reduces performance on Latent Hatred. Further case analysis suggests that entity-level polarity can serve as a useful weak retrieval signal, but it is still too coarse to determine whether a hate-related concept is being endorsed, criticized, quoted, or merely mentioned.

Overall, this study demonstrates that an attribute-enriched knowledge graph can provide useful external evidence for implicit hate speech detection. By organizing hate-related implications into structured triplets and enriching them with semantic and affective attributes, HateKG helps the model reason beyond surface-level textual cues and improves both the reasoning ability and interpretability of implicit hate speech detection.
URI: http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/106052
DOI: 10.6342/NTU202602501
全文授權: 未授權
電子全文公開日期: N/A
顯示於系所單位:資訊管理學系

文件中的檔案:
檔案 大小格式 
ntu-115-1.pdf
  未授權公開取用
630.27 kBAdobe PDF
顯示文件完整紀錄


系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。

社群連結
聯絡資訊
10617臺北市大安區羅斯福路四段1號
No.1 Sec.4, Roosevelt Rd., Taipei, Taiwan, R.O.C. 106
Tel: (02)33662353
Email: ntuetds@ntu.edu.tw
意見箱
相關連結
館藏目錄
國內圖書館整合查詢 MetaCat
臺大學術典藏 NTU Scholars
臺大圖書館數位典藏館
本站聲明
© NTU Library All Rights Reserved