﻿<?xml version="1.0" encoding="utf-8"?>
<ArticleSet>
  <ARTICLE>
    <Journal>
      <PublisherName>مرکز منطقه ای اطلاع رسانی علوم و فناوری</PublisherName>
      <JournalTitle>Journal of Information Systems and Telecommunication (JIST) </JournalTitle>
      <ISSN>2322-1437</ISSN>
      <Volume>14</Volume>
      <Issue>55</Issue>
      <PubDate PubStatus="epublish">
        <Year>2026</Year>
        <Month>11</Month>
        <Day>9</Day>
      </PubDate>
    </Journal>
    <ArticleTitle>Hybrid Embedding and GPT-5 Instruction-Augmented Framework for Arabic Hate Speech Detection</ArticleTitle>
    <VernacularTitle>Hybrid Embedding and GPT-5 Instruction-Augmented Framework for Arabic Hate Speech Detection</VernacularTitle>
    <FirstPage>268</FirstPage>
    <LastPage>276</LastPage>
    <ELocationID EIdType="doi">10.66224/jist.51935.14.55.268</ELocationID>
    <Language>en</Language>
    <AuthorList>
      <Author>
        <FirstName> Anas</FirstName>
        <LastName> Khudhur Abbas al juboori</LastName>
        <Affiliation></Affiliation>
      </Author>
      <Author>
        <FirstName>Heshaam</FirstName>
        <LastName>Faili</LastName>
        <Affiliation>Tehran</Affiliation>
      </Author>
    </AuthorList>
    <History PubStatus="received">
      <Year>2025</Year>
      <Month>10</Month>
      <Day>28</Day>
    </History>
    <Abstract>&lt;p class="Sammary" style="page-break-after: auto;"&gt;Hate speech is defined as language that creates hatred, hostility, or violence towards people or groups based on their religion, ethnicity, gender, and/or nationality. As social media continues to grow and expand, the influx of hate speech and abusive language on these platforms is on the rise and has become a significant threat to online safety and social cohesion. Despite improvements in the detection of hate speech with automation technologies, there is little research conducted on the detection of hate speech in Arabic, especially regarding varying dialects. Current approaches often fail to identify language patterns or contextual nuances that are specific to different dialects. Moreover, little research has been conducted focused solely on instruction-based learning in this field. We presented a methodology for detecting abusive language and hate speech in Arabic utilizing the L-HSAB database. Through the combination of AraBERT-v2 and Multilingual MPNet, hybrid embeddings were created that have both language-specific characteristics and cross-linguistic conceptual characteristics. UMAP and HDBSCAN were applied for both dimension reduction and cluster analysis, and iterative refinement of clustering was facilitated by the utilization of Silhouette and Coherence scores. Based on the identified resulting clusters generated, GPT-5 was used to produce linguistically grounded instructional prompts, which were then assessed by both LLM-as-a-Judge using prompt generation and by human validators. Finally, the identified instructions were incorporated into further fine-tuning of the models. Experimental results indicate that the proposed method yields 90.5% accuracy for the purpose of detecting hate speech in dialects of Arabic.&lt;/p&gt;</Abstract>
    <ObjectList>
      <Object Type="Keyword">
        <Param Name="Value">Hate Speech</Param>
      </Object>
      <Object Type="Keyword">
        <Param Name="Value">Large Language Model</Param>
      </Object>
      <Object Type="Keyword">
        <Param Name="Value">Supervised Fine Tuning</Param>
      </Object>
      <Object Type="Keyword">
        <Param Name="Value">Arabic</Param>
      </Object>
    </ObjectList>
    <ArchiveCopySource DocType="Pdf">http://jist.ir/ar/Article/Download/51935</ArchiveCopySource>
  </ARTICLE>
</ArticleSet>