Jul 2026· International Journal for Research in Applied Science and Engineering Technology· Vol 14, pp. 1972-1979· 0 citations
TL;DR
This study explores a low-resource approach to detecting hate speech in English and Swahili code-switched text by fine-tuning pre-trained language models, and shows that fine-tuning modern language models can offer a practical and scalable solution for hate speech detection in multilingual environments.
Abstract
The use of social media in East Africa has grown rapidly, and with it, the spread of hate speech has become a serious
concern. This problem is even more complex in online spaces where people often switch between English and Swahili within the
same sentence or conversation. Such code-switching makes it difficult for existing systems to accurately detect harmful content,
especially because there is limited labeled data and much of the language used is informal and context-dependent. This study
explores a low-resource approach to detecting hate speech in English and Swahili code-switched text by fine-tuning pre-trained
language models. In this work, transformer-based models such as BERT and AfriBERTa are adapted to better understand
mixed-language communication. The models are trained on a carefully prepared dataset made up of real social media posts that
reflect how people actually write and speak online. These posts are manually labeled to capture both direct and subtle forms of
hate speech, including expressions that are influenced by local culture and everyday slang. The findings show that fine-tuned
models perform better than traditional machine learning approaches, especially in terms of accuracy and overall detection
quality. They are also more effective at handling informal language, abbreviations, and mixed grammar structures. Beyond
performance, the study also looks at fairness and bias, emphasizing the need for systems that are sensitive to cultural and
linguistic diversity. Overall, this work shows that fine-tuning modern language models can offer a practical and scalable
solution for hate speech detection in multilingual environments.
It is challenging to detect hate speech in Low Resource Languages (LRLs) because of the absence of annotated data, the informality of its language structure, and the lack of standardized grammar. A good example of such a challenge is Roman Urdu which is broadly used by South Asians on social media and has a high variation while lacking contextually consistent spellings. The objective of this paper is to conduct a comprehensive assessment of Large Language Models (LLMs) for Hate Speech Detection (HSD) in Roman Urdu script and fine-tune these models using the Parameter-Efficient Fine-Tuning (PEFT) method called Low-Rank Adaptation (LoRA). To evaluate zero-shot inference, we benchmarked it against PEFT on different transformer models, including Mistral, LLaMA, Falcon, and multilingual BERT. Experiments are conducted on the PURUTT (Parallel Urdu and Roman Urdu Corpus for Toxic Comments and Transliteration) dataset with over 72,000 annotated comments. The results suggest that zero shot models perform moderately (F1 = 0.56), but updating a small fraction of the model trainable parameters improves the classification performance significantly (F1>0.93). Our results have shown that PEFT delivers outstanding performance alongside excellent computational efficiency, making it highly suitable for low-resource language processing tasks.
Toneema Zubair, Muhammad Asif, F. Kamiran et al.· 0 citations
The rapid expansion of social media has accelerated the spread of hate speech, particularly within multilingual and code-switched environments where users frequently alternate languages within a single conversation. Detecting this kind of content is still challenging because of the use of multiple languages, the lack of clear context, and the difficulty of identifying the intended target of the hate speech. Current methods for detecting hate speech, such as traditional machine learning models and transformer-based architectures such as Bidirectional Encoder Representations from Transformers (BERT) and Cross-lingual Language Model-RoBERTa (XLM-R), have improved contextual understanding. However, they still struggle to accurately identify the intended target of hate speech and lack fine-grained context awareness. This limitation reduces the effectiveness of implicit hate speech detection and results in more false positives, especially in multilingual and low-resource settings. To tackle these issues, this research introduces a Context-Aware and Target-Adaptive Multilingual Hate Speech Detection (CTM-HSD) model. The proposed method combines multilingual transformer-based embeddings with a context-aware attention mechanism to capture semantic dependencies in text. It also includes a separate target identification module to identify the person or group being targeted. In addition, adaptive learning techniques, such as transfer learning and data augmentation, are used to improve performance in low-resource and code-switched scenarios. Evaluation on multilingual and code-switched datasets shows that the proposed model outperforms baseline models such as BERT and XLM-R, with an accuracy of 92.4% and an F1-score of 91.2%. The findings demonstrate that the integration of contextual awareness and target adaptability markedly enhances the identification of implicit hate speech and reduces false positives. The proposed framework offers a robust and flexible solution for real-world multilingual content moderation systems.
K. Shruthi, K. Shivanna· Engineering, Technology &...· 0 citations
Hate speech on digital communication platforms has become a major obstacle to healthy online discourse, especially in multilingual societies such as India, where Hindi is a dominant language in social media interactions. Hostile, offensive, and defamatory speech is linguistically and socio-culturally complex because of colloquial idioms, regional variations, code-mixing, and culturally embedded references. However, effective detection is crucial for creating safer and more inclusive digital communication environments. This study evaluates advanced language models for analysing hate speech in Hindi social media content. A dataset of 21000 Hindi posts from Twitter and public repositories, categorized into five categories, was analysed: hate, offensive, fake, defamation, and non-hostile. General-purpose models were tested against Hindi-specific language models, including Hindi-BERT (based on Bidirectional Encoder Representations from Transformers, or BERT) and MuRIL, to investigate whether performance can be further enhanced by integrating Long Short-Term Memory (LSTM) layers. In tests, the best-performing model, Hindi-BERT (with sequential learning added), correctly identified 94 out of every 100 posts—a considerable improvement over simpler models. For social media platforms, it has real-world consequences: it can automatically identify harmful content to be reviewed by a human, minimize nuisance alerts that waste human moderators' time, and identify defamatory or bogus posts before they gain a lot of traction. The results offer a systematic approach to researching how hostility develops in the Hindi-speaking online community, how linguistic creativity (e.g., slang, sarcasm, code-mixing) can conceal or manifest hostility, and how decisions about content moderation influence public discourse for communication scholars. Overall, this paper illustrates that language-specific computational tools can be used for both platform governance and communication research, provided that the cultural context is considered. Finally, technical methods are combined with communication scholarship to explain and curb harmful speech in the online public sphere.
Rachna Narula, Vedika Gupta, Jawad Khan et al.· Journal of Communication, La...· 0 citations
Hate speech that occurred in social media platforms presents critical safety challenges, especially in language environments with minimal source where text is often written in an informal form, abbreviated, and context dependent. This paper further provides empirical evidence regarding the ability of the BERT (Bidirectional Encoder Representations from Transformers) model in detecting hate speech in a rough text environment and comparing directly with other deep learning techniques.Three Bidirectional Long Short–Term Memory (BiLSTM) variants using Word2Vec, FastText, and TF-IDF representations were compared with BERT, which employs contextual language representations. The data is separated by using a train-test ratio of 80:20, and the performance is evaluated according to their accuracy, precision, recall, F1-score, and area under the curve (AUC). The result shows that BERT outperforms all variations of BiLSTM by achieving an accuracy of 87.18%, an F1-score of 87.08%, and an AUC matrix of 0.9244. According to the efficiency matters, BiLSTM and Term Frequency–Inverse Document Frequency (TF-IDF) are the most ineffective for their uneven classification distribution, while BiLSTM with Word2Vec and FastText shows moderate effectivity. This finding conclusively demonstrates the benefit of transformer-based models to capture the subtleties of language with noisy textual data, which is commonly found in low-resource settings. To conclude the above discussion, this finding suggests that BERT has great potential as a multilingual content moderation tool to be applied in informal and unstructured digital environments.
Received: 25 August 2025 | Revised: 29 January 2026 | Accepted: 23 June 2026
Conflicts of Interest
The authors declare that they have no conflicts of interest to this work.
Data Availability Statement
The data that support the findings of this study are openly available in GitHub at https://github.com/okkyibrohim/id-multi-label-hate-speech-and-abusive-language-detection, reference number [29].
Author Contribution Statement
Yosia Immanuel Bastian: Conceptualization, Methodology, Formal analysis, Resources, Data curation. Aditiya Hermawan: Software, Supervision, Project administration, Writing – original draft, Writing – review & editing, Visualization. Ardiane Rossi Kurniawan Maranto: Validation, Investigation. Benny Daniawan: Validation, Investigation. Junaedi Junaedi: Validation, Investigation.
Yosia Immanuel Bastian, Aditiya Hermawan, Ardiane Rossi Kurniawan Maranto et al.· Artificial Intelligence and...· 0 citations
Detecting hate speech in low-resource and unseen languages remains challenging due to limited labeled data and linguistic diversity. This paper presents a comparative study of zero-shot cross-lingual transfer for hate speech detection using two multilingual transformer models: mDeBERTa-v3 and XLM-RoBERTa. To the best of our knowledge, mDeBERTa-v3 has not been previously used by researchers for zero-shot cross-lingual hate speech detection, making this the first study to evaluate its capabilities in this task. Furthermore, we introduce new unseen languages that have not been studied before in this context, including Hebrew, Amharic, and Swahili, alongside other languages such as Indonesian, Danish Portuguese, Turkish, French, and Russian. We evaluate model performance under three training scenarios: a single source language (Turkish), semantically similar language clusters, and multiple clusters from different language families. Experimental results show that mDeBERTa-v3 consistently outperforms XLM-RoBERTa in zero-shot settings. The most notable improvement is observed for Hebrew, where the macro F1 score increases from 0.39 (XLM-RoBERTa) to 0.71 (mDeBERTa-v3), a gain of 0.32. Substantial gains are also seen for Amharic (0.52 → 0.73, +0.21), Indonesian (0.57 → 0.71, +0.14), and Swahili (0.65 → 0.75, +0.10). Across all experimental conditions, mDeBERTa-v3 achieves average macro F1 gains ranging from 0.04 to 0.19, with statistical significance (p < 0.02). The model’s advantage is attributed to its disentangled attention mechanism, which enables better generalization across typologically distant languages. These findings establish mDeBERTa-v3 as a novel and more robust architecture for zero-shot cross-lingual hate speech detection, particularly for previously unexplored low-resource languages.
Ghadeer Al-Badani, M. Alsurori, Akram Alsubari· 2026 6th International Confe...· 0 citations
Social media platforms such as X, formerly Twitter, have become major channels for communication, information sharing and public discussion. However, the rapid growth of user-generated content has also increased the spread of hate speech and offensive language. Automated hate speech detection remains challenging in multilingual and code-mixed environments, where users frequently combine languages, informal spelling, slang, abbreviations and culturally specific expressions. In Malaysia, online discourse often involves Malay-English code-mixing, commonly referred to as Manglish, which creates additional challenges for natural language processing systems. This study presents an evidence-informed systematic review and research-readiness taxonomy for culturally aware Malay-English hate speech detection. Unlike conventional reviews that mainly summarize model architecture and performance scores, this review critically evaluates existing studies based on dataset availability, code-mix authenticity, annotation practice, cultural sensitivity, model architecture, evaluation strategy, explainability, robustness and deployment readiness. To strengthen this review, this article incorporates a completed empirical case study on Manglish hate-speech detection, using posts collected from X (formerly Twitter). The case study used keyword-based data collection, Malaya NLP-based language filtering, bilingual manual annotation, manual class balancing and transformer-based evaluation using mBERT, XLNet and XLM-RoBERTa. The case evidence is used only as an empirical lens to illustrate practical challenges in dataset curation, class imbalance, lexical overlap and model robustness. It is not positioned as a new benchmark dataset or an independent experimental contribution. The review finds that existing Malay and Malay-English resources remain fragmented. Some datasets are monolingual Malay hate speech datasets, some are bilingual but language-separated, while others are code-mixed but developed for sentiment analysis rather than hate speech detection. Transformer-based models such as BERT, mBERT and XLM-RoBERTa show strong potential, but their results are difficult to compare due to inconsistent datasets, label definitions, class distributions, evaluation metrics and limited robustness testing. The main novelty of this review is the proposed taxon- omy that evaluates the field through six readiness dimensions, which are data, linguistic, cultural, modelling, evaluation and deployment readiness. The findings from this study provide a structured foundation for developing culturally aware, robust, explainable and deployable hate speech detection systems for Malay-English code-mixed social media.
F. Azmi, Normaisharah Mamat, Rawad Abdulghafor et al.· IEEE Access· 0 citations