@inproceedings{reddy-renjit-2026-hate,
    title = "Does Hate Transfer? Cross-Lingual Generalisation of Offensive Content Detection Across {I}ndic Languages",
    author = "Reddy, Purandhar M.  and
      Renjit, Sara",
    editor = "Mitkov, Ruslan  and
      Mu{\~n}oz, Rafael  and
      Lloret, Elena  and
      Ranasinghe, Tharindu  and
      Estevanell-Valladares, Ernesto L.  and
      Lamsiyah, Salima  and
      Montoyo, Andr{\'e}s  and
      Ezzini, Saad",
    booktitle = "Proceedings of the Second International Conference on Natural Language Processing and Artificial Intelligence for Cyber Security",
    month = jun,
    year = "2026",
    address = "Alicante, Spain",
    publisher = "Department of Languages and Information Systems, University of Alicante",
    url = "https://aclanthology.org/2026.nlpaics-1.24/",
    pages = "228--231",
    abstract = "A common assumption in low-resource NLP is that cross-lingual transfer from a related language can substitute for target-language annotation when labelled data is scarce. We test this assumption for offensive content detection across five Indic languages by evaluating all twenty directed transfer pairs from a LLaMA3.1-8B model fine-tuned with Low-Rank Adaptation (LoRA) on the MACD benchmark. Only three of twenty pairs achieve tolerable transfer loss below 15{\%}, all involving Malayalam as the source language. Telugu is the hardest transfer target (average loss 33.8{\%}), while Malayalam is the most transferable source (average loss 16.8{\%}). Confusion-matrix analysis reveals two distinct failure modes: Tamiland Kannada-trained models are conservative under-flaggers that miss 73{--}82{\%} of offensive content with near-zero false alarms, while Malayalam-trained models are aggressive flaggers that miss far less (39{\%}) but over-flag at 21{\%}. These patterns do not follow typological structure: a Spearman correlation between URIEL typological similarity and transfer F1 yields {\ensuremath{\rho}} = {\ensuremath{-}}0.254 (p = 0.281), failing to conf irm the typological hypothesis. Our results indicate that cross-lingual shortcuts are unreliable for this task and that language-specific annotation cannot be avoided by appealing to linguistic family membership."
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="reddy-renjit-2026-hate">
    <titleInfo>
        <title>Does Hate Transfer? Cross-Lingual Generalisation of Offensive Content Detection Across Indic Languages</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Purandhar</namePart>
        <namePart type="given">M</namePart>
        <namePart type="family">Reddy</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Sara</namePart>
        <namePart type="family">Renjit</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-06</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the Second International Conference on Natural Language Processing and Artificial Intelligence for Cyber Security</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Ruslan</namePart>
            <namePart type="family">Mitkov</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Rafael</namePart>
            <namePart type="family">Muñoz</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Elena</namePart>
            <namePart type="family">Lloret</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Tharindu</namePart>
            <namePart type="family">Ranasinghe</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Ernesto</namePart>
            <namePart type="given">L</namePart>
            <namePart type="family">Estevanell-Valladares</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Salima</namePart>
            <namePart type="family">Lamsiyah</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Andrés</namePart>
            <namePart type="family">Montoyo</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Saad</namePart>
            <namePart type="family">Ezzini</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Department of Languages and Information Systems, University of Alicante</publisher>
            <place>
                <placeTerm type="text">Alicante, Spain</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
    </relatedItem>
    <abstract>A common assumption in low-resource NLP is that cross-lingual transfer from a related language can substitute for target-language annotation when labelled data is scarce. We test this assumption for offensive content detection across five Indic languages by evaluating all twenty directed transfer pairs from a LLaMA3.1-8B model fine-tuned with Low-Rank Adaptation (LoRA) on the MACD benchmark. Only three of twenty pairs achieve tolerable transfer loss below 15%, all involving Malayalam as the source language. Telugu is the hardest transfer target (average loss 33.8%), while Malayalam is the most transferable source (average loss 16.8%). Confusion-matrix analysis reveals two distinct failure modes: Tamiland Kannada-trained models are conservative under-flaggers that miss 73–82% of offensive content with near-zero false alarms, while Malayalam-trained models are aggressive flaggers that miss far less (39%) but over-flag at 21%. These patterns do not follow typological structure: a Spearman correlation between URIEL typological similarity and transfer F1 yields \ensuremathρ = \ensuremath-0.254 (p = 0.281), failing to conf irm the typological hypothesis. Our results indicate that cross-lingual shortcuts are unreliable for this task and that language-specific annotation cannot be avoided by appealing to linguistic family membership.</abstract>
    <identifier type="citekey">reddy-renjit-2026-hate</identifier>
    <location>
        <url>https://aclanthology.org/2026.nlpaics-1.24/</url>
    </location>
    <part>
        <date>2026-06</date>
        <extent unit="page">
            <start>228</start>
            <end>231</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Does Hate Transfer? Cross-Lingual Generalisation of Offensive Content Detection Across Indic Languages
%A Reddy, Purandhar M.
%A Renjit, Sara
%Y Mitkov, Ruslan
%Y Muñoz, Rafael
%Y Lloret, Elena
%Y Ranasinghe, Tharindu
%Y Estevanell-Valladares, Ernesto L.
%Y Lamsiyah, Salima
%Y Montoyo, Andrés
%Y Ezzini, Saad
%S Proceedings of the Second International Conference on Natural Language Processing and Artificial Intelligence for Cyber Security
%D 2026
%8 June
%I Department of Languages and Information Systems, University of Alicante
%C Alicante, Spain
%F reddy-renjit-2026-hate
%X A common assumption in low-resource NLP is that cross-lingual transfer from a related language can substitute for target-language annotation when labelled data is scarce. We test this assumption for offensive content detection across five Indic languages by evaluating all twenty directed transfer pairs from a LLaMA3.1-8B model fine-tuned with Low-Rank Adaptation (LoRA) on the MACD benchmark. Only three of twenty pairs achieve tolerable transfer loss below 15%, all involving Malayalam as the source language. Telugu is the hardest transfer target (average loss 33.8%), while Malayalam is the most transferable source (average loss 16.8%). Confusion-matrix analysis reveals two distinct failure modes: Tamiland Kannada-trained models are conservative under-flaggers that miss 73–82% of offensive content with near-zero false alarms, while Malayalam-trained models are aggressive flaggers that miss far less (39%) but over-flag at 21%. These patterns do not follow typological structure: a Spearman correlation between URIEL typological similarity and transfer F1 yields \ensuremathρ = \ensuremath-0.254 (p = 0.281), failing to conf irm the typological hypothesis. Our results indicate that cross-lingual shortcuts are unreliable for this task and that language-specific annotation cannot be avoided by appealing to linguistic family membership.
%U https://aclanthology.org/2026.nlpaics-1.24/
%P 228-231
Markdown (Informal)

[Does Hate Transfer? Cross-Lingual Generalisation of Offensive Content Detection Across Indic Languages](https://aclanthology.org/2026.nlpaics-1.24/) (Reddy & Renjit, NLPAICS 2026)

ACL