Beyond words: a hybrid transformer-ensemble approach for detecting hate speech and offensive language on social media

Machine learning based approaches

Early machine learning (ML) approaches demonstrated strong performance on explicit hate speech. Greevy & Smeaton (2004) tackled the issue of identifying racist text using supervised ML techniques. The research involved transforming textual data into numerical vectors through the extraction of bigram features, supplementary to other methods such as bag-of-words (BOW) and part-of-speech (POS) tagging. To analyze the outcomes of their experiments, they used the SVM classifier. They documented a high accuracy rate of 91.50% when employing the polynomial kernel function alongside BOW representations.

Building on these foundations, Davidson et al. (2017) employed a combination of unigrams, bigrams, and trigrams, each weighted by TF-IDF, to assess the content of tweets sourced via a crowdsourced hate speech lexicon. The tweets were divided into three categories (hate speech, offensive language, or neither), leveraging the crowdsourcing method for categorization. To handle the complexity of data, they utilized L1 and L2 regularization techniques and applied LR and support vector machines (SVM) to train their models. They obtained a precision, recall, and F1-score of 91%, 90%, and 90%, respectively.

More recently, Khanday et al. (2022) classified hate speech on Twitter during the COVID-19 era by employing TF-IDF, BOW, and Tweet length engineering techniques for feature extraction. The authors fed the generated data to different ML algorithms as well as applied different ensemble learning techniques. According to their findings, the SGB classifier exceeds all others with precision, recall, F1-score, and accuracy of 98.04%, 99%, and 97%, respectively. Although these ML methods achieve high accuracy, they depend on manual feature engineering and struggle with contextual nuances like sarcas, which require domain-specific tuning for optimal performance.

Deep learning based approaches

Later work leveraged the learned representations to reduce feature engineering. Gambäck & Sikdar (2017) explored a variety of methods for tweet classification, incorporating character 4-grams, semantic-based word vectors, random word vectors, and a combination of word vectors and character n-grams. They implemented max pooling to simplify the features before applying a softmax function for the final tweet classification. Their research concluded that the model incorporating word2vec embeddings was superior in terms of performance, achieving an F1-score of 78.3% validated by a 10-fold cross-validation process. Extending these techniques to multilingual contexts, Ghosh, Ghosh & Das (2017) carried out sentiment identification using code-mixed text data obtained from social media. The author experimented using English-Bengali and English-Hindi code-mixed datasets. Data were categorized as positive, negative, or neutral based on the polarity of the statement. They achieved a polarity classification accuracy of 68.5% by utilizing a multilayer perception model in conjunction with SentiWordNet, opinion lexicon, and POS tags.

For sequential modeling, Pitsilis, Ramampiaro & Langseth (2018) developed an ensemble method using recurrent neural networks (RNN) with LSTM architecture to analyze a dataset of roughly 16,000 publicly available tweets. They transformed tweets into vector form through word frequency vectorization and included additional features to gauge the likelihood of users posting hate speech. To enhance the classifier’s accuracy, the outputs from multiple LSTM classifiers were aggregated, resulting in a notable F1-score of 0.9320. Ishmam & Sharmin (2019) delved into the classification of a Bengali language comments dataset obtained from Facebook, using both ML algorithms and a deep neural network based on the gated recurrent unit (GRU) structure. The features were extracted using n-gram and character n-gram methodologies, each with their own TF-IDF values, after a series of data preprocessing steps. They fed the processed data to several ML algorithms and initially achieved an accuracy of 52.20%. This was further improved by using a GRU-based model, which increased the accuracy by approximately 18%, reaching 70%.

A significant research study in 2019 focused on creating a Danish dataset for the detection of hate speech and abusive language (Sigurbergsson & Derczynski, 2019). The dataset employed by the authors comprised comments from both Reddit and Facebook, along with details regarding the types and target audiences of offensive language. To achieve the highest F1-score of 0.74, the authors utilized DL models together with diverse feature sets. Abro et al. (2020) proposed an automatic scheme to detect hate speech messages. The author used various significant ML algorithms with different feature engineering techniques and compared the performance of the models. Their experimental results demonstrated that the SVM exhibited the highest accuracy of 79% when bigram features were applied.

Aljero & Dimililer (2021) proposed a novel approach to detect hate speech in English tweets by using a stacked ensemble architecture. Their method combined three classifiers, namely SVM, LR, and XGB trained on word2vec and universal encoding features. To generate the final output, they made the meta classifier LR to integrate the predictions from the base classifiers along with their corresponding features. The experimental outcomes of their proposed architecture demonstrated a performance improvement across the four datasets (HatEval, Davidson (Dvd), COVID-HATE, and ZeerakW) compared to standard stacking, base classifiers, and majority voting approaches. Although DL reduced feature engineering needs, these approaches still required large labeled datasets and struggled with computational costs, particularly for low-resource languages and real-time applications.

Transformer learning based approaches

Transformers revolutionized multilingual detection. Biradar, Saumya & Chauhan (2022) conducted a study in 2022 to detect hate speech in Hinglish. The study explored the efficacy of transformer models such as IndicBERT and multilingual Bidirectional Encoder Representation (mBERT), as well as transfer learning techniques utilizing pre-trained language models such as ULMFiT, and BERT. In addition, they proposed a deep neural network feature extraction model and a transformer-based interpreter (TIF-DNN). According to their experimental findings, their suggested model works better than current cutting-edge techniques for identifying hate speech in Hinglish with an accuracy of 73%.

Malik et al. (2024) conducted a comprehensive comparative study evaluating both traditional ML and DL based approaches for hate speech detection using three widely adopted benchmark datasets: Davidson, Founta, and Twitter Sentiment Analysis. Their work systematically assessed various embedding techniques, including TF-IDF, GloVe, and transformer-based models like BERT, Small BERT, ALBERT, and ELECTRA. Classifiers ranging from SVM and XGBoost to CNN and bidirectional long short-term memory (Bi-LSTM) were analyzed for effectiveness, efficiency, and domain generalization, providing a robust baseline for future transformer-ensemble integrations.

In the context of identifying hate speech on Twitter, text normalization has been underexplored, particularly its impact on handling lexical variants and enhancing model performance. Mansur et al. (2024) has proposed an improved normalization technique that combines rule-based patterns and the SymSpell spelling correction algorithm to convert out-of-vocabulary (OOV) words into recognized vocabulary, thus improving the effectiveness of sentiment and hate speech detection tasks. Although transformers reduce the dependency on text normalization, work like Mansur et al. (2024) demonstrates that careful OOV handling (via SymSpell+rules) can still boost performance by 12%, suggesting opportunities for hybrid approaches.

In a recent advancement, researchers explored the adaptation of five cutting-edge transformer models, ALBERT, DeBERTa, ELECTRA, HateBERT and DeepSeek, for the task of cyberbullying detection in social media contexts (Philipo et al., 2025). These models, originally developed for general text classification or sentiment analysis, were refined to address the nuanced and often implicit nature of harmful content online. The study highlighted how these modern architectures can effectively capture emotional tone and contextual cues, surpassing the limitations of traditional keyword-based methods. This underscores the growing importance of sentiment-aware transformer models in enhancing the sensitivity and contextual understanding required for robust hate speech and cyberbullying detection.

Research on online hate speech is hampered by limited resources, mainly affecting languages with limited data sets. These barriers are often due to the high cost and difficulty in collecting and managing training data in such cases. Recent advances in areas such as text (Raffel et al., 2020; Brown et al., 2020), image (Ramesh et al., 2022; Saharia et al., 2022), and speech generation (Shen et al., 2018) have opened up avenues for new analytical approaches. This enhancement enables the generation of synthetic data from model output, which in turn can be used to retrain new models. Data augmentation methods demonstrated their effectiveness in building a hate speech classifier (Khullar et al., 2024; Ansari, Kaur & Saxena, 2024; Githa et al., 2024), and demonstrated impressive performance. Additionally, some studies have addressed resource scarcity by developing sentiment corpora in low-resource languages, such as Malay, using semi-supervised learning methods (Sukawai & Omar, 2020).

Although multilingual studies have expanded to languages like German (Velankar, Patil & Joshi, 2022), Dutch (Markov, Gevers & Daelemans, 2022), and English (del Valle-Cano et al., 2023), most solutions face a trilemma: (1) ML/DL models struggle with contextual nuances, (2) pure transformer approaches overlook class imbalance, and (3) hybrid methods lack systematic optimization of transformer-ensemble synergy.

Our work addresses these gaps by introducing a novel hybrid architecture that combines RoBERTa-large’s contextual depth with XGBoost’s imbalance robustness, advancing beyond existing hybrid approaches (Biradar, Saumya & Chauhan, 2022) through rigorous hyperparameter tuning often overlooked in prior ensemble studies (Aljero & Dimililer, 2021). The proposed model demonstrates practical viability by achieving state-of-the-art performance across benchmark datasets (Davidson, HateSpeechDetection, ToxicTweets) while avoiding the extensive data requirements characteristic of pure DL methods (Pitsilis, Ramampiaro & Langseth, 2018), offering an optimized balance between transformer-based contextual understanding and ensemble learning’s classification robustness.

Datasets

Kaggle is one of the largest data source providers for learning purposes, so for this study we used a publicly available dataset of toxic tweets from Kaggle, namely the ToxicTweets (TT) dataset (Iyer, 2021). We prepare it appropriately to achieve our objectives. The dataset approximately includes 56,744 tweets. The dataset has a target variable, namely toxicity. We also evaluated our model on two other datasets that are also publicly available. One dataset is from literature, namely the Davidson dataset (Davidson et al., 2017) and the other is a HateSpeechDetection (HSD) (https://doi.org/10.6084/m9.figshare.19686954.v1) dataset (Cooke, 2022) which is a combination of comments from different platforms such as Twitter, Reddit, 4Chan, etc. The description of the datasets is given in Table 2.

The TT dataset has two classes, namely toxic which is labeled as 1 and non-toxic which is labeled as 0. Davidson’s dataset has three classes; hateful, offensive, and neither; while the HSD dataset also has two classes; hateful and not hateful labeled 1 and 0, respectively. As far as our knowledge extends, the TT and HSD datasets are novel contributions to the field of hate speech detection, since no prior studies have been conducted on these two datasets. Therefore, our study introduces these datasets to the field for the first time. The visualization of datasets through a pie chart to see the percentage of each class is given in Fig. 5.

In previous research, the main issue researchers faced was labeling of the data, specifically determining which text is hateful and which is not. A significant challenge in this task is the difficulty in identifying hate speech, especially when it is disguised through sarcasm or lacks explicit words associated with hatred, stereotyping, or racism (Mansur, Omar & Tiun, 2023). The main cause of this issue is that a proper definition of hate speech was not available at that time. Over time, many researchers have proposed definitions of hate speech, which serve as the basis for labeling datasets today. Although the datasets used in this study were pre-labeled, we present examples of tweets from these datasets in Table 3 to illustrate instances of both hateful and non-hateful content. The presence of specific keywords in the text often serves as the basis for categorizing tweets. Table 3 helps provide a clearer understanding of the nature and structure of the data.

Preprocessing

Data manipulation or deletion before usage can be referred to as data preprocessing, which is done to ensure or improve performance. In this section, we discuss the steps that we used to clean the data to ensure that it performs well when a model is evaluated on it and generates better results.

• Handling the diacritics in the text: To handle the diacritics in the text, we used Python’s “unicodedata” library, which allows standard character encoding. As we know, text in English is not readable for computers, so encoding using Unicode is necessary.

• Replacement of Twitter handles and URLs: We replaced all Twitter handles (i.e., usernames starting with ‘@’) and URLs (i.e., any strings starting with ‘http’ or ‘https’) in tweets with an empty string.

• Emoji removal: We removed emojis from the text, as they do not contribute to the model’s understanding of hate speech or offensive language.

• Contractions expansion: To preserve the full meaning of contractions (e.g., “don’t” to “do not”), we used the Python ‘contractions’ library to expand them, as expanded text improves interpretability for the model.

• Lowercasing text: All text was converted to lowercase to ensure consistency, as models often benefit from treating words in a case-insensitive manner.

• Tokenization and Lemmatization: We used the ‘WordNetLemmatizer’ from NLTK to lemmatize words after tokenizing the tweets. Lemmatization reduces words to their base or root form, making the data more consistent and reducing redundancy.

• Elimination of stop words: Stop words, which are frequently used words with little semantic meaning, were removed to make training more efficient. Common words such as “and,” “the,” “is,” etc., were removed using NLTK’s stop word list.

• Removal of special characters, digits, and short words: We removed all non-alphabetic characters, digits, and words with a length of 1 or 2 characters (such as “&”, “#”, “1”, “4”) as they do not contribute meaningfully to the detection of hate speech.

• Concatenation of words: Finally, the tokens of each tweet were concatenated into a single string, with each word separated by spaces. This is useful for further processing or analysis that requires data in a text string format.

Feature extraction using RoBERTa-Large

In this study, we used the pre-trained RoBERTa-Large model for feature extraction. RoBERTa is a transformer-based model that has shown substantial improvements in various NLP tasks due to its ability to generate contextualized embeddings. For traditional ML models, we employed TF-IDF as the feature extraction method, which represents text based on term frequency and inverse document frequency. Instead of relying on these frequency-based features, RoBERTa captures complex linguistic patterns by encoding each word in relation to the context of surrounding words.

During training, each tweet is passed through the RoBERTa-Large model to generate high-dimensional embeddings, which are then pooled to create a single feature representation for each text. This pooled embedding serves as the input feature for our classifier, significantly enhancing the model’s ability to understand nuanced language, including hate speech and offensive language. The complete feature extraction pipeline is detailed in Algorithm 1.

Furthermore, data were segregated, with 80% designated for the training set and the remaining 20% for the testing set. Table 4 presents the distribution of data among the training and testing sets for all three datasets. This split is fundamental for a thorough assessment of the models, allowing for confident and trustworthy findings from both trained and novel test data.

Algorithmic description

ML is a field of study that focuses on developing algorithms that enable computers to learn from observations and interactions in the real world. These algorithms allow machines to simulate human-like behavior and enhance their ability to learn and perform well with input data (Madaan et al., 2021). ML employs a range of techniques to deal with data issues (Mahesh, 2020). Data scientists typically emphasize that there is no single “best” type of algorithm that is suitable for all situations. The choice of algorithm depends on various factors, such as the problem being addressed, the number of variables involved, and the type of model that will yield the best results. Other factors, such as available data and computing resources, also play a role in selecting the most appropriate algorithm for a given task.

In this study, we evaluated several ML algorithms in comparison to our proposed model. The following subsections provide an overview of each algorithm, including our new approach, RoBERTa-Large + XGBoost, which significantly improves the detection of hate speech and offensive language. Furthermore, Fig. 6 illustrates the key components and their interconnections in the proposed algorithm/model, providing a visual representation of the methodology employed in this study.

Model evaluation metrics

Model evaluation is a technique used to analyze the effectiveness of a model based on some constraints (Sheikh, Goel & Kumar, 2020). However, while evaluating the model, it should be kept in mind that it cannot overfit or step on the model. To measure the performance of a model, there are several evaluation metrics available such as confusion metrics, accuracy, precision, recall, F1-score etc.

In binary classification, accuracy is measured concerning both positive and negative outcomes using the following formula:

(8) $$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$where TP = true positives, TN = true negatives, FP = false positives, and FN = false negatives.

The precision value, indicating the accuracy of our model, is given as follows:

(9) $$Precision=\frac{TP}{TP+FP}.$$

Recall (TPR) measures the proportion of actual positive instances that are correctly identified by the model. A high recall indicates that the model is good at identifying positive instances.

(10) $$Recall=\frac{TP}{TP+FN}.$$

F1-score is the harmonic mean (HM) of the precision and recall scores.

$$F1\text{-score}=\frac{2}{\frac{1}{Precision}+\frac{1}{Recall}}$$

(11) $$F1\text{-score}=\frac{2\times Precision\times Recall}{Precision+Recall}.$$

Experimental setup

The architecture of the hybrid model consisted of a RoBERTa-Large backbone (355M parameters) modified with an added dropout layer (p = 0.3) and a linear classification head. The transformer component was fine-tuned using AdamW optimization (learning rate $=2\times {10}^{-5}$, ${\beta }_1=0.9$, ${\beta }_2=0.999$) with a batch size of 18 and a sequence length capped at 130 tokens. Training ran for five epochs with cross-entropy loss, using a linear learning rate scheduler without warm-up steps. The XGBoost classifier was trained on 1,024-dimensional RoBERTa embeddings with the following configuration: learning rate = 0.01, n-estimators = 400, max-depth = 4, and multi-class log loss objective.

All experiments were conducted on dual NVIDIA RTX 3080 GPUs (10 GB VRAM each) with PyTorch 1.10 and XGBoost 1.5, utilizing full FP32 precision. The implementation leveraged HuggingFace Transformers for RoBERTa integration and scikit-learn for evaluation metrics. Hyperparameters such as learning rate, batch size, and number of epochs were manually tuned based on validation set performance. We found that training for fewer than five epochs resulted in underfitting, while training for more epochs led to overfitting. To mitigate overfitting, a dropout rate of 0.3 was applied, and validation accuracy was closely monitored throughout training.

Results

While we experimented with a total of nine ML models, we specifically focused on discussing and highlighting the performance of the models that achieved superior results. Tables 5, 6 and 7 display the performance of each classifier, including SVC, RFC, BC, AdaBoost, LGBM, and XGBoost trained to detect hate speech and offensive language across the TT, Davidson, and HSD datasets. The results highlight the comparative advantage of our RoBERTa-Large + XGBoost approach in terms of accuracy and contextual understanding.

The results on the TT dataset, shown in Table 5, are presented in terms of accuracy, precision, recall, and F1-score for each classifier. The PA, utilizing RoBERTa-Large embeddings with an XGBoost classifier, outperformed all other models, achieving the highest accuracy of 0.92 and an F1-score of 0.91. This indicates that the PA model provides a more balanced performance in detecting hate speech and offensive language with improved precision and recall as well. Among the other classifiers, both LGBM and traditional XGBoost achieved high accuracy scores of 0.91, with corresponding F1-scores of 0.90. Although SVC showed slightly better recall (0.91) than the other models, its overall performance metrics did not surpass those of the PA model. This suggests that RoBERTa-Large embeddings contribute additional contextual understanding that enhances the model’s ability to accurately classify nuanced language in tweets, giving it an edge over traditional models using TF-IDF-based features.

Similarly, the results obtained on the Davidson dataset, presented in Table 6, show that the PA, combining RoBERTa-Large embeddings with XGBoost, achieved the highest performance across all metrics. PA recorded an accuracy of 0.92 and an F1-score of 0.91, outperforming all other models. Both LGBM and traditional XGBoost followed closely with accuracy scores of 0.91, although they fell short in precision and recall. The superior performance of PA is likely due to RoBERTa-Large’s ability to capture contextual nuances in language, which enhances classification accuracy when paired with XGBoost’s robust handling of complex feature relationships. This combination allows for more precise detection of hate speech and offensive language compared to traditional TF-IDF based models.

At last, Table 7 shows that the PA achieved the highest scores on the HSD dataset, with an accuracy of 0.97 and an F1-score of 0.96. RFC followed with strong results across all metrics at 0.95, while Adaboost and Bagging also performed well, each scoring 0.94 across metrics. XGBoost performed slightly lower, achieving an accuracy of 0.93.

Limitations and future work

This study has certain limitations that should be acknowledged:

Language-specific focus: The scope of this research is restricted to English-language tweets, which limits its applicability to other languages or multilingual social media platforms.

Data bias and ethics: The datasets used in this study, including two relatively new datasets (TT and HSD), may not fully capture the diversity of hateful content across different platforms, regions, or time periods. As pre-annotated datasets, they are also subject to annotation subjectivity and cultural biases, which can affect model fairness and reliability. The novelty of TT and HSD means that their annotation quality and bias profiles are less well-established than those of benchmark datasets like Davidson. Finally, deploying automated hate speech detection systems carries risks of misclassification, which require careful consideration in real-world use.

Modality limitation: The model is designed to detect hate speech only from textual data, which excludes multimodal forms of hate speech commonly shared on social media, such as hateful images, videos, or memes.

Interpretability consideration: While the PA achieves strong predictive performance, the current study does not incorporate post-hoc explainability tools such as SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) or attention visualizations. Including such tools in future work could further enhance the transparency and trustworthiness of the model’s decisions, particularly in sensitive application areas.

Building on identified limitations, future research could broaden the scope by incorporating multimodal data, such as images and videos alongside text, to enable a more comprehensive analysis of hate speech on social media platforms. Additionally, future studies could explore hate speech detection in multiple languages beyond English, such as Hindi, Urdu, and Japanese, to broaden the applicability of these techniques. An important future direction is the inclusion of emojis, which often convey emotional or contextual cues in online communication. We plan to develop an emoji-to-text conversion module and investigate how emoji semantics interact with hate speech content, potentially improving detection performance. We also plan to focus on models that can more effectively address class imbalance issues and perform well on multiclass data, as these are key challenges in hate speech detection.

Further, to enhance transparency and accountability, we aim to integrate interpretability techniques into our future work. This includes applying SHAP and LIME for post-hoc inspection, as well as attention visualization mechanisms inherent in transformer architectures. These tools will help elucidate which features or tokens influence classification decisions and ensure the model’s behavior aligns with human-understandable reasoning. Given the strong performance of hybrid models, we also aim to leverage this approach for other classification problems, combining model strengths to yield better results than individual models alone. This study is intended as a guide for newcomers to the field, laying out the complete process of text classification using ML and Transformer-based approaches in hate speech detection.