The social media boom in multilingual communities has resulted in the proliferation of ofending and harmful content, particularly in low-resourced Dravidian languages, where the overall phenomenon of code-mixing with English is widespread. TriFusion is a multi-model pipeline to detect ofensive language in Kannada, Tulu, Tamil, and Malayalam. The TriFusion system combines TFIDF with Ensemble classifiers, IndicBERT with MLP, and XLM-RoBERTa with XGBoost, with further improvement of majority-vote fusion. In the DravidianCodeMix 2025 shared task, the proposed system ranked Rank 3rd in Malayalam (macro F1-Score = 0.750), 5th in Tulu (macro F1-Score = 0.730), 7th in Kannada (macro F1-Score = 0.375), and 7th in Tamil (macro F1-Score = 0.369). On the whole, the findings indicate that transformers can be generalized well across languages, and Ensemble models are more efective under low-resource settings.
Automated moderation systems have been required by the propagation of abusive and hateful speech
on social media platforms [
The DravidianCodeMix@FIRE 2025 shared task on Ofensive Language Detection, which aims to detect
ofensive content in code-mixed YouTube comments, is the source of the datasets used in this study
[15][
Language Tamil Malayalam Kannada Tulu
Train
The proposed ofensive language detection system has been illustrated in Figure 1, which incorporates preprocessing, feature extraction, classification, and majority voting.
To prepare the code-mixed datasets for classification of ofensive language, a thorough preprocessing pipeline was implemented to reduce noise and normalize inputs. The following are the stages: • Noise reduction: Unnecessary elements that are commonly included in social media comments, including links, mentions, hashtags, and irregular symbols, are eliminated. We kept only English and language-specific characters and eliminated URLs, user mentions, hashtags, digits, and special characters. • Emoji Management: The function demojize() converted emojis into their descriptive text equivalents. To preserve sentiment cues, this semantic information was retained. Emoji descriptors were then surrounded by colons, which were removed to simplify the token structure. • Unicode normalization: With an emphasis on language-specific characters, we performed
Unicode normalization using the IndicNormalizerFactory from the indic-nlp-library.
The proposed system integrates a soft-voting ensemble of six traditional machine learning classifiers [16], including LR, XGB [17], MLP, RF, Support Vector Classifier(SVC), and KNN. The score of each model is a probability distribution over the target classes, which is averaged to select the most probable class based on the mean probability, therefore the ensemble can take advantage of individual model confidence and generalize across models better than through hard voting. Lowercasing, removing punctuations as well as normalizing code-mixing content are done on the input text, which then get TFIDF vectorization [6] and the limit of features was set to 50000 with the n-gram range of (1, 2). The system training and testing use each language’s data separately, class labels are encoded via LabelEncoder, and the results are assessed by using Accuracy, macro F1-score ( mF1-score), and weighted F1-score (wF1-score). The resulting sparse representations are the inputs to the ensemble, which makes the final prediction.
The IndicBERT , a transformer multilingual model, which is optimized on Indian languages, to create dense contextual embedding of the input text. we fine-tune the representations of the [CLS] token of the last hidden layer based on the maximum sequence length of 128. These embeddings are fed into a pre-trained MLP classifier, which consists of two hidden layers having 256 and 128 nodes, respectively, an activation function of the form ReLU, and an optimizer of the form Adam, trained with up to 300 iterations. The system is trained on each language dataset respectively using LabelEncoder to encode labels and scored against the development set using Precision, Recall and macro F1-score. Extending the capacity of IndicBERT [18] to support rich semantic and syntactic features and that of MLP to estimate non-linear decision boundaries, multi-lingual classification can be performed robustly using this fusion technique.
The XLM-RoBERTa [19] transformer model that our system uses is pre-trained on more than 100 languages, including some Indian languages, which is favorable to multilingual and code-mixed scenarios. Each input sentence has its embedding as [CLS] token (768-dimensional) extracted at the final hidden layer, and with maximum length up to 128 tokens. Then, these embeddings get passed to an XGBoost classifier that is set to multi-class classification and that has a maximum depth of 6,200 estimators, and a learning rate of 0.1. To overcome the impact of imbalance between classes, we calculate weights of classes inversely proportional to the frequencies and use them in training. The system is optimized on development set based on the usual classification accuracy indices and the final predictions on test set are produced through mapping between the model output and the original label set. The strategy would blend the strong contextual ability of XLM-R with the gradient boosting functionality of XGBoost to be able to deal efectively with both high-dimensional embeddings and imbalanced data.
To take advantage of the complementary skill sets of our individual models, this study chose a majority voting ensemble approach. Namely, the predictions of three diferent systems were combined: (i) an ensemble of a classical set of classifiers trained on the basis of TFIDF. (ii) an IndicBERT model with a MLP classifier. (iii) an XLM-RoBERTa model with an XGBoost classifier.
The three systems were queried in each test cases and the predicted label of the three systems per test case were used to obtain the final output label by adopting hard (majority) voting. Where no label had a majority then the mode function within the Pandas library was used to pick the most common label. The purpose of this ensemble technique is to reduce the bias of deep contextual embeddings and increase the overall robustness of approximations of the true model and achieve better overall performance in classification.The result is shown in Table 3
The proposed models were tested on four Dravidian languages i.e. Tulu, Malayalam, Kannada and Tamil using three important evaluation measures such as Accuracy (2), mF1-score, and wF1-score. Such parameters combine two aspects of the prediction correctness overall and the capability of the model to manage the aspect of the imbalance of classes, which plays a significant role in ofensive language detection. The findings show that transformer-based architectures are always superior to conventional ensemble models in most languages, especially in Malayalam and Tamil. These two languages had the best scores in the XLM-RoBERTa model with an mF1-score of 0.76 and 0.73 in this order, indicating that the model has a good grasp of the context of the code-mixed text. In the case of Tulu, though, TFIDF + Ensemble model made a competitive mark of 0.73 on mF1-score, meaning that classical statistical features have not yet been discarded in small or less diverse datasets. However, Kannada was more challenging with high data imbalance and morphological complexity, with TFIDF a little more successful than deep models on all three metrics. TriFusion ensemble, a sergeant that is composed of the results of TFIDF, IndicBERT, and XLM-RoBERTa based on majority voting, exhibited better language-independent generalization. It scored the best in Accuracy in Malayalam (0.92), and all mF1 and wF1-scores of all languages were constant, which narrowed the performance diference between the resource-rich and low-resource cases. In general, the comparison between these three parameters depicts that transformerbased models perform better in the case of well-represented datasets, whereas hybrid ensemble fusion ofers a more robust and balanced result in detecting multilingual and code-mixed ofensive language.
TF-IDF + Ensemble
IndicBERT + MLP
XLM-R + XGBoost Tulu Malayalam Kannada Tamil
Accuracy
The error analysis was used to analyze the diferences in performance between models, with the main emphasis on misclassifications in the minority classes. The comparison of Accuracy, mF1-score, and wF1-Score was found to show that the Accuracy was good across the board, but low mF1-Scores were a manifestation of the way the models treated underrepresented classes like Ofensive Targeted and Ofensive Non-Targeted. As an example, Tamil and Kannada got a decent Accuracy (0.77 and 0.70) but significantly lower mF1-Score (0.37 and 0.39), indicating the issues of class imbalance.
The XLM-RoBERTa model was found to be able to capture contextual nuances and classify nonofensive categories, as well as generally overfit to the most common non-ofensive classification types, but IndicBERT was unable to scale to fine-grained diferences because it did not support code-mixed text representation. TriFusion ensemble alleviated such problems by taking advantage of model diversity, providing a higher balance between classes, especially in Malayalam (mF1-score 0.41). The remaining errors were due to the ambiguous nature of cases on sarcasm, transliterated slang, and mixed sentiment polarity where context on interpretation was very important. On balance, the findings indicate that although hybrid models such as TriFusion can raise the level of robustness, more necessary steps must be implemented to reduce bias via data augmentation strategies, context-enriched embeddings, and contrastive learning to improve the detection of minority classes.
The scripts of the proposed methodology, including the preprocessing scripts, feature extraction, and model training modules, are publicly available on Github Repository.
This paper introduced TriFusion, a multi-model ensemble that used ofensive language detection in code-mixed Dravidian languages, namely, Kannada, Tulu, Tamil, and Malayalam. The system showed that in low-resource and linguistically diverse conditions, the combination of lexical and contextual features can help to boost the robustness of the traditional machine learning classifiers with the aid of transformer-based deep representations. The experiments also showed that the contextual cues are well captured by the transformer architectures like XLM-RoBERTa, and that the imbalance limitation typical of underrepresented languages is addressed by the hybrid fusion. TriFusion performed well on all languages, especially on Malayalam, and this proved that ensemble fusion is less biased compared to single models.
To overcome the drawbacks associated with minority class detection as noted in the present study, in the next work, suggested to apply the method of data augmentation and context-enriched embedding. Contrastive learning and cross-lingual alignment may also be used together to increase the resilience of diferentiating subtle ofensive material in underrepresented groups. These guidelines will assist in moving the field of ofensive language detection on low-resource, code-mixed Dravidian languages to a more fair and generalized state.
In the course of preparing this manuscript, the author(s) employed the generative AI tool ChatGPT. Its use was limited to performing checks for grammar and spelling. Following this, the author(s) conducted a thorough review and revision of the text and assume full responsibility for the final published content. in dravidian code-mixed languages, in: Forum of Information Retrieval and Evaluation FIRE-2025, 2025. [6] R. Adyanthaya, R. Shetty, Yenlp_cs@ dravidianlangtech 2025: Sentiment analysis on code-mixed tamil-tulu data using machine learning and deep learning models, in: Proceedings of the Fifth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages, 2025, pp. 288–292. [7] T. Davidson, D. Warmsley, M. Macy, I. Weber, Automated hate speech detection and the problem of ofensive language, in: Proceedings of the international AAAI conference on web and social media, volume 11, 2017, pp. 512–515. [8] R. Kumar, G. Bhanodai, R. Pamula, M. R. Chennuru, Trac-1 shared task on aggression identification: Iit (ism)@ coling’18, in: Proceedings of the first workshop on trolling, aggression and cyberbullying (TRAC-2018), 2018, pp. 58–65. [9] Z. Talat, D. Hovy, Hateful symbols or hateful people? predictive features for hate speech detection on twitter, in: Proceedings of the NAACL student research workshop, 2016, pp. 88–93. [10] T. Mandl, S. Modha, A. Kumar M, B. R. Chakravarthi, Overview of the hasoc track at fire 2020: Hate speech and ofensive language identification in tamil, malayalam, hindi, english and german, in: Proceedings of the 12th annual meeting of the forum for information retrieval evaluation, 2020, pp. 29–32. [11] B. R. Chakravarthi, R. Priyadharshini, N. Jose, T. Mandl, P. K. Kumaresan, R. Ponnusamy, J. P.
McCrae, E. Sherly, et al., Findings of the shared task on ofensive language identification in tamil, malayalam, and kannada, in: Proceedings of the first workshop on speech and language technologies for Dravidian languages, 2021, pp. 133–145. [12] A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. Guzmán, E. Grave, M. Ott, L. Zettlemoyer, V. Stoyanov, Unsupervised cross-lingual representation learning at scale, arXiv preprint arXiv:1911.02116 (2019). [13] D. Kakwani, A. Kunchukuttan, S. Golla, G. NC, A. Bhattacharyya, M. M. Khapra, P. Kumar, Indicnlpsuite: Monolingual corpora, evaluation benchmarks and pre-trained multilingual language models for indian languages, in: Findings of the association for computational linguistics: EMNLP 2020, 2020, pp. 4948–4961. [14] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, in: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), 2019, pp. 4171–4186. [15] B. R. Chakravarthi, M. B. Jagadeeshan, V. Palanikumar, R. Priyadharshini, Ofensive language identification in dravidian languages using mpnet and cnn, International Journal of Information Management Data Insights 3 (2023) 100151. URL: https://www.sciencedirect.com/science/article/ pii/S2667096822000945. doi:https://doi.org/10.1016/j.jjimei.2022.100151. [16] J. Preetham, J. Anitha, Ofensive language detection in social media using ensemble techniques, in: 2023 International Conference on Circuit Power and Computing Technologies (ICCPCT), IEEE, 2023, pp. 805–808. [17] R. Razzak, S. T. O. Taki, M. R. Mim, M. S. H. Patwary, M. I. Pavel, Vulgar comments classification:
Comparison between cnn, xgboost and svm (2023). [18] A. Garain, A. Mandal, S. K. Naskar, Junlp@ dravidianlangtech-eacl2021: Ofensive language identification in dravidian langauges, in: Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Languages, 2021, pp. 319–322. [19] M. Xu, S. Liu, Rb_bg_mha: A roberta-based model with bi-gru and multi-head attention for chinese ofensive language detection in social media, Applied Sciences 13 (2023) 11000.