2.1 Ensemble learning

Ensemble learning aims to generate a set of classifiers (i.e. ensemble members or base models), and then merge their outputs to obtain the final prediction (Dietterich Reference Dietterich2000). There are various ensemble strategies (Ganaie et al. Reference Ganaie, Hu, Malik, Tanveer and Suganthan2022), including conventional classifier ensembles (Opitz and Maclin Reference Opitz and Maclin1999) such as Bagging (Breiman Reference Breiman1996) and Boosting (Schapire Reference Schapire1990). Other methods include negative correlation learning (Liu and Yao Reference Liu and Yao1999), ensemble pruning (Toraman and Can Reference Toraman and Can2012), and explicit/implicit methods such as Dropconnect (Wan et al. Reference Wan, Zeiler, Zhang, LeCun and Fergus2013) and Dropout (Srivastava et al. Reference Srivastava, Hinton, Krizhevsky, Sutskever and Salakhutdinov2014).

There are different techniques for ensemble construction. Stacking aims to construct multiple classifiers with different model types. On the other hand, other ensemble approaches employ the same model by learning different data subsets, for example Bagging (Breiman Reference Breiman1996) employs a decision tree algorithm on different data subsets. In this study, we follow the latter approach, that is using the same model (i.e. homogeneous ensemble generation) but learning different data subsets, to exploit different characteristics of hate speech datasets.

The output of ensemble members is merged in fusion strategies, for example unweighted model averaging, majority voting, and super learner (Laan, Polley, and Hubbard Reference Laan, Polley and Hubbard2007). In this study, we employ unweighted model averaging as our fusion strategy, and Transformer-based language models (Devlin et al. Reference Devlin, Chang, Lee and Toutanova2019; Sanh et al. Reference Sanh, Debut, Chaumond and Wolf2019) as backbone classifier.

2.2 Text classification and hate speech detection

Automated hate speech detection is widely studied in the natural language processing community (Schmidt and Wiegand Reference Schmidt and Wiegand2017). Text content is useful for extracting linguistic and syntactical features with the bag-of-words (Nobata et al. Reference Nobata, Tetreault, Thomas, Mehdad and Chang2016; Waseem Reference Waseem2016; Burnap and Williams Reference Burnap and Williams2016; Davidson et al. Reference Davidson, Warmsley, Macy and Weber2017). User-based features, including meta-attributes and user profiles, can be good signals to detect hate speech (Waseem Reference Waseem2016; Chatzakou et al. Reference Chatzakou, Kourtellis, Blackburn, Cristofaro, Stringhini and Vakali2017; Unsvåg and Gambäck Reference Unsvåg and Gambäck2018). Zhang et al. (Reference Zhang, Wojatzki, Horsmann and Zesch2019) study feature engineering, specifically by adding lexico-semantic features and showing the improvements by various architectures such as CNNs, LSTMs, and SVM-based classifiers. Word embeddings (Pennington, Socher, and Manning Reference Pennington, Socher and Manning2014; Nobata et al. Reference Nobata, Tetreault, Thomas, Mehdad and Chang2016; Mou, Ye, and Lee Reference Mou, Ye and Lee2020) and deep neural networks (Hochreiter and Schmidhuber Reference Hochreiter and Schmidhuber1997; Kim Reference Kim2014) are also employed to represent text semantics. Recently, BERT (Devlin et al. Reference Devlin, Chang, Lee and Toutanova2019) and similar encoder models based on Transformer (Vaswani et al. Reference Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser and Polosukhin2017) outperform previous methods for hate speech detection (Liu et al. Reference Liu, Ott, Goyal, Du, Joshi, Chen, Levy, Lewis, Zettlemoyer and Stoyanov2019; Tekiroğlu, Chung, and Guerini Reference Tekiroğlu, Chung and Guerini2020; Kennedy et al. Reference Kennedy, Jin, Mostafazadeh Davani, Dehghani and Ren2020; Caselli et al. Reference Caselli, Basile, Mitrović and Granitzer2021; Mathew et al. Reference Mathew, Saha, Yimam, Biemann, Goyal and Mukherjee2021; Röttger et al. Reference Röttger, Vidgen, Nguyen, Waseem, Margetts and Pierrehumbert2021; Toraman et al. Reference Toraman, Sahinuç and Yilmaz2022a).

Multi-task learning is leveraged by integrating hate speech detection with emotion detection to reflect the potential correlations between hate speech and certain negative emotion states (Min et al. Reference Min, Lin, Li, Zhao, Lu, Yang and Xu2023). Supervised contrastive learning is utilized for capturing span-level information for hate speech detection (Lu et al. Reference Lu, Lin, Zhang, Li, Zhang, Zong, Ma and Xu2023). The role of social stereotypes in hate speech detection is examined by the impact of social stereotypes on annotation and classification algorithms (Davani et al. Reference Davani, Atari, Kennedy and Dehghani2023).

Furthermore, there are other efforts such as extracting hate spans from hateful tweets using Transformer-based language models (Khan, Ma, and Vosoughi Reference Khan, Ma and Vosoughi2021; Zhou et al. Reference Zhou, Caines, Pete and Hutchings2022), and dedicated shared tasks such as Toxic Spans Detection (Pavlopoulos et al. Reference Pavlopoulos, Sorensen, Laugier and Androutsopoulos2021; Ranasinghe and Zampieri Reference Ranasinghe and Zampieri2021).

Instead of detecting hateful text segments, PAN21 is a shared task that profiles hate speech spreaders (Bevendorff et al. Reference Bevendorff, Chulvi, Sarracén, Kestemont, Manjavacas, Markov, Mayerl, Potthast, Rangel, Rosso, Stamatatos, Stein, Wiegmann, Wolska and Zangerle2021). The dataset in this task consists of multiple tweets of the user. The participants have solutions to detect the overall tendency for spreading hate speech including Convolutional Neural Networks (CNN) (Siino et al. Reference Siino, Di Nuovo, Tinnirello and La Cascia2021). In this paper, we detect the existence of hate speech instead of hate speech spreaders. Yet, we utilize some author features such as the number of followers and followees in our ensemble methods.

2.3 Ensemble learning for hate speech detection

Ensemble learning is applied for hate speech detection. There are efforts to combine text features an ensemble of different classifiers with a voting scheme (Anusha and Shashirekha Reference Anusha and Shashirekha2020; Hegde, Anusha, and Shashirekha Reference Hegde, Anusha and Shashirekha2021; Mutanga et al. Reference Mutanga, Naicker and Olugbara2022). Alsafari, Sadaoui, and Mouhoub (Reference Alsafari, Sadaoui and Mouhoub2020) create the ensemble of CNN and BiLSTM classifiers with contextual and non-contextual word-embedding models. Zimmerman, Kruschwitz, and Fox (Reference Zimmerman, Kruschwitz and Fox2018) examine the ensemble of three CNN-based models whose weights are initialized randomly and trained on the same dataset. Tula et al. (Reference Tula, Potluri, Ms, Doddapaneni, Sahu, Sukumaran and Patwa2021) utilize an ensemble of different models such as distilmBERT (Sanh et al. Reference Sanh, Debut, Chaumond and Wolf2019) and ULMFiT (Howard and Ruder Reference Howard and Ruder2018), and loss functions. Agarwal and Chowdary (Reference Agarwal and Chowdary2021) propose an ensemble learning-based adaptive model that works towards overcoming the strong user bias present in the dataset.

Recent advances in language models provided by the Transformer architecture (Vaswani et al. Reference Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser and Polosukhin2017) are also utilized with ensemble learning. Turban and Kruschwitz (Reference Turban and Kruschwitz2022) find that Transformer models are well-suited for ensemble learning by data augmentation techniques. Moreover, some participants in the Hate Speech Detection (HaSpeeDe) shared task at EVALITA (Sanguinetti et al. Reference Sanguinetti, Comandini, Nuovo, Frenda, Stranisci, Bosco, Caselli, Patti and Russo2020) use ensemble techniques such as fine-tuning different Transformer-based language models on the same dataset, and then create an ensemble of these models to improve their results. Similarly, there are efforts to utilize an ensemble of different Transformer-based language models for Arabic hate speech detection (Mubarak, Al-Khalifa, and Al-Thubaity Reference Mubarak, Al-Khalifa and Al-Thubaity2022; Magnossão de Paula et al. Reference Magnossão de Paula, Rosso, Bensalem and Zaghouani2022).

There are several other shared tasks such as HASOC 2021 (Mandl et al. Reference Mandl, Modha, Shahi, Madhu, Satapara, Majumder, Schäfer, Ranasinghe, Zampieri, Nandini and Jaiswal2021) and OffensEval 2019 (Zampieri et al. Reference Zampieri, Malmasi, Nakov, Rosenthal, Farra and Kumar2019) and most of the participants propose methods based on ensemble learning. Contrary to our study, ensemble methods in such studies propose combining different classifiers and model architectures (Nikolov and Radivchev Reference Nikolov and Radivchev2019; Kumar, Roy, and Saumya Reference Kumar, Roy and Saumya2021).

Our main difference from the studies that utilize ensemble learning for hate speech detection is that we construct our ensembles by using different aspects of hate speech, such as topics and users. Our focus is not on the classification algorithm or voting scheme of ensemble learning but on the way of constructing ensemble members or base models by data splitting. That is, we investigate the effect of different aspects of the datasets by keeping the model architecture the same and modifying the approach of training and ensemble construction. One can employ different classification models for each split of the dataset, however, the improvement by ensemble construction might be affected by the choice of models. Moreover, we conduct detailed experiments to understand the performance of the proposed ensemble methods for hate speech detection in multiple languages.

3.1 Random ensemble

As a baseline ensemble approach, we apply the bagging method (Breiman Reference Breiman1996), which randomly divides the original train set into 80% of the train and 20% of the validation set. The same process is repeated five times to obtain five ensemble members. We use the split ratio of 80-20 to have consistent training size with the main approach illustrated in Fig. 2. The best model is selected based on the validation loss for each member. We call this approach Random Ensemble since the instances in data splits are obtained randomly. The random ensemble is similar to the Combination methods in terms of train size.

3.2 Fold ensemble

Another baseline approach is the fold ensemble. In this method, the training set is stratified and divided into five folds. That is, each fold contains 20% of the train data, and there is no overlap among the fold splits. To construct an ensemble member, four of five data splits are merged (80% of the train data). This process is repeated five times with different four splits (i.e. three of five data splits). The remaining 20% of data is used for validation. In contrast to Random Ensemble, we ensure that the ratio of the classes is preserved, and each ensemble member has a different train set with 75% of overlap with other ensemble members. The fold ensemble is similar to the Combination method in terms of train size.

3.3 Topic ensemble

Random and fold ensembles do not necessarily represent different aspects of hate speech. In topic ensemble, the train set is divided according to specific hate topics such as race and gender. We rely on existing topic labels in the datasets. If topic labels do not exist, we filter the dataset by topic keywordsFootnote ^b to obtain topic sets. That is, tweets are assigned to a particular topic label by the existence of a topic keyword. An instance can be assigned to multiple topics.

An example of topic splitting is illustrated in Fig. 3a. In this figure, we divide a dataset used in our experiments to show an example of how topic splits are represented in UMAP embedding space (McInnes and Healy Reference McInnes and Healy2018) using DistilBERT embeddings (Sanh et al. Reference Sanh, Debut, Chaumond and Wolf2019). Although the instances belonging to particular topics are mainly clustered in the embeddings space, we observe no strict distinction among different topics, represented by different colors in the figure. This observation might show that there can be overlapping parts across topics (e.g. gender-based discussions in religions), and the Topic ensemble method can extract different aspects of hate speech exposed in different topics. Since topics are content features, we refer to the Topic ensemble method as a content-based ensemble method. Validation and test splits are generated as in the main approach illustrated in Fig. 2.

Figure 3. Data splits for different ensemble methods (Topic, Influential, and GMM) in the Toraman22-EN dataset by using UMAP representation (McInnes and Healy Reference McInnes and Healy2018). Best viewed in color.

3.4 Influential ensemble

The user’s writing style can be an important factor in hate speech detection. We assume that users interacting with more people tend to have different writing styles (Martin, Sintsova, and Pu Reference Martin, Sintsova and Pu2014). Such influential users can have different writing styles due to their popularity. Based on the tweet author’s influential score, we divide the training set into five splits with an equal number of samples in each. We use the influential score, similar to (Toraman et al. Reference Toraman, Sahinuç, Yilmaz and Akkaya2022b), as $ \log{(n_{follower}/n_{followee})}$ where $ n_{follower}$ and $ n_{followee}$ are the numbers of followers and followees, respectively. The tweets that are shared by the users are then used in training.

An example of influential splitting is illustrated in Figure 3b. We observe significant overlapping parts across different influential splits, represented by different colors in the figure. This observation might suggest that the difference in writing styles of different influential groups is not based on the context of tweets, and the Influential Ensemble method can extract different writing styles of tweet authors who have different user features, particularly being influential in this case. We refer to the Influential ensemble method as a user-based ensemble method. Validation and test splits are generated as in the main approach illustrated in Fig. 2.

3.5 GMM ensemble

Topic ensemble relies on topic labels or keywords. To extract different topics of hate speech in an unsupervised way without labels and keywords, we use Gaussian Mixture Modeling (GMM) (Reynolds Reference Reynolds2009) that clusters the training set as a mixture of five Gaussian distributions. We choose Gaussian Mixture Modeling as the clustering method in this study for the following reasons. GMM can describe a complex set with a mixture of Gaussian distributions, which can fit the task of text clustering with pre-trained embeddings (Aharoni and Goldberg Reference Aharoni and Goldberg2020). Moreover, we use pre-trained embeddings as feature vectors, and GMM outperforms other clustering methods, such as the k-means clustering (Lloyd Reference Lloyd1982), when pre-trained embeddings are employed for representing topics in text (Sia, Dalmia, and Mielke Reference Sia, Dalmia and Mielke2020).

We use GMM to assign a sample to the most probable distribution to apply hard assignments. The GMM is initialized randomly using a fixed seed. An example of GMM splitting is illustrated in Fig. 3c. We observe a clear distinction of clusters or data splits, represented by different colors in the figure. The GMM ensemble can be useful for understanding the performance of hate ensembles when the contents in data splits are closely similar to each other. Since GMM is a probabilistic model that assumes that the instances are obtained from a mix of Gaussian distributions based on the input contents, we refer to the GMM ensemble method as a content-based ensemble method.

3.6 Combination ensemble

In contrast to using data folds as individual ensemble members, we use each fold as a validation set and combine the remaining data as a training set in the Combination method (i.e. 80% of data is for training and 20% of data is for validation), as represented in Fig. 2d. When the Combination method is not applied, training is based on a unique 20% of data, since there are five folds and we select a different fold each time for training. However, the Combination method has overlapping instances among five ensemble members since it merges four folds each time. The motivation is to increase the number of instances to be used in the training of each ensemble member while combining different aspects of hate speech.

We apply the Combination method to the Topic, Influential, and GMM methods. Note that the baseline methods, Random and Fold, have already 80% of data in the train. In other words, the normal versions of the Topic, Influential, and GMM methods have smaller training sizes than the baselines, which may cause a disadvantage in training. However, they can still challenge baselines as observed in the Experiments section.

4.1 Datasets

There are five datasets used in the experiments. Davidson17 (Davidson et al. Reference Davidson, Warmsley, Macy and Weber2017) and Founta18 (Founta et al. Reference Founta, Djouvas, Chatzakou, Leontiadis, Blackburn, Stringhini, Vakali, Sirivianos and Kourtellis2018) are two widely used datasets for the task of hate speech detection. Toraman22-EN (Toraman et al. Reference Toraman, Sahinuç and Yilmaz2022a) is a recent large-scale dataset with topic information, which is an important factor in creating ensemble members in this study. In order to generalize the results to multiple languages, we also use the datasets in Italian (IHSC) (Sanguinetti et al. Reference Sanguinetti, Poletto, Bosco, Patti and Stranisci2018) and Turkish (Toraman22-TR) (Toraman et al. Reference Toraman, Sahinuç and Yilmaz2022a). The reason for selecting these languages is that our ensemble approach requires topic and user information, which are partly provided in these datasets. There are also other datasets for hate speech detection such as SOLID (Rosenthal et al. Reference Rosenthal, Atanasova, Karadzhov, Zampieri and Nakov2021) with 9 million samples annotated in a self-supervised fashion. However, in this study, we rely on manually annotated datasets.

We query Twitter API to get additional features for the tweet authors in Founta18, Toraman22, and IHSC such as the number of followers and accounts being followed, to utilize them in Influential Ensemble. Moreover, since only the Toraman22 datasets provide topic labels, we filter the other datasets to obtain the instances related to topics. The final distribution of class labels and topics used in the experiments are given for all datasets in Tables 1 and 2, respectively. The details of how we obtain the dataset versions used in the experiments are given in the following subsections for each dataset.

Table 1. Class distributions of the datasets used in the experiments

* The dataset includes no sample in the Spam class.

** The IHSC dataset includes no sample in the Offensive or Abusive class.

Table 2. Topic distributions of the datasets used in the experiments

* Some instances have multiple topic labels.

** The dataset includes no sample in the Disability topic.

4.2 Experimental design

4.2.3 Evaluation metric

We use stratified 5-fold validation. For each fold, 20% of the data is set aside as the test set, while the remaining 80% is further split into a train set and validation set. Since the number of ensemble members is set to five, we split this 80% of data into five parts of equal size, and then obtain train and validation sets based on the ensemble methods explained in Section 3. The validation sets are used for finding the optimal models according to their validation losses during a number of epochs of training. Within each fold, we train five base models with the same backbone using different train-validation pairs derived from the 80% part of the whole data as explained in Section 3. The softmax probabilities of five models are unweighted-averaged to get the final prediction.

We evaluate models with the F1 score using scikit-learn (Pedregosa et al. Reference Pedregosa, Varoquaux, Gramfort, Michel, Thirion, Grisel, Blondel, Prettenhofer, Weiss, Dubourg, Vanderplas, Passos, Cournapeau, Brucher, Perrot and Duchesnay2011). The F1 score is calculated as follows.

\begin{equation*} F1 = 2 \times \frac{Precision \times Recall}{Precision + Recall} \end{equation*}

where $Precision$ is the ratio of the correct predictions over all predictions, and $Recall$ is the ratio of the correct predictions over all true samples. We apply the weighted F1 score to get the average score over all classes since most of the datasets in the experiments are imbalanced.

We report the average of 5-fold results, along with their standard deviation. Moreover, we apply a paired t-test to find any statistically significant difference at a 95% interval in pairwise comparisons between two mean scores.

4.2.4 Baselines

Our focus in this study is the way of constructing ensemble members or base models by data splitting. We, therefore, compare our ensemble construction methods with the following baseline methods:

• • Single-Text: We fine-tune a single model with the same size of train data. The motivation is to understand the performance when no ensemble method is applied.

• • Single-Text-Topic: We fine-tune a single classifier that includes the text features and topic information. CLS text embeddings from the Transformer-based language models are concatenated with topic embedding (one-hot vector) that represents the tweet’s topic. This concatenated embedding vector is then given to a linear classification layer.

• • Single-Text-Influential: We fine-tune a single classifier that includes the text features and user features. CLS text embeddings are concatenated with influential features, which are follower and followee counts. The features are then given to a linear classification layer.

• • Ensemble-Bagging: A baseline ensemble construction method is to create ensemble members by selecting data randomly, as explained in Section 3.

• • Ensemble-Fold: A baseline ensemble construction method is to create ensemble members by splitting data discretely, as explained in Section 3.

We refer to our ensemble construction methods (i.e. Topic, GMM, Influential, and their Combination versions) as Ensemble-MethodName in the experimental results. For instance, Ensemble-Topic refers to the ensemble approach explained in Section 3.3.

Table 3. Comparison of ensemble methods for hate speech detection. An average of 5-fold is reported in terms of the F1 score. The highest scores for each class and dataset are given in bold. The symbol “^*” indicates a statistically significant difference using a paired t-test at a 95% interval in pairwise comparisons between the highest-performing method and others

4.3 Experimental results

We present the results of ensemble methods for multi-class hate speech detection in all datasets in Table 3. We omit the results of a class label when the dataset includes no instances for this class label. Our observations that reflect the answers to the research questions of this study are given as follows.

RQ-1: Improvements by ensemble learning. We examine if ensemble learning provides any improvements to the performance of hate speech detection. Therefore, we pair the performance scores of a single model with an ensemble model. For instance, there are two single and six ensemble models in Davidson17, resulting in 12 pairs concerning Weighted F1 score. We observe that the highest-performing method in terms of the weighted F1 score is a type of ensemble learning, compared to the Single classifier, in all datasets. The ensemble methods yield better scores than the single methods in 67 of 108 (62%) comparisons in terms of the weighted F1 score. Furthermore, the Combination ensemble methods have better results than the single methods in 31 of 40 (78%) comparisons. The Combination method merges different splits to increase the number of instances and thereby improves the performance in most cases.

RQ-2: Benefits of ensemble methods that leverage different aspects of hate speech. When we compare our ensemble methods with the baseline ensemble approaches, we find that our proposed ensemble construction methods perform the highest in terms of the F1 score of the Hate class in four of five datasets (80% of the datasets) and in terms of the F1 score of the Offensive class in three of four datasets (75% of the datasets). We thereby argue that our ensemble construction methods can capture better different aspects of hate speech compared to the baseline ensemble construction methods. On the other hand, the Fold method performs the highest for the Hate class and weighted F1 score in Toraman22-EN. The reason could be the balanced distribution of both classes and topics in this dataset. Similarly, the Fold method performs highest in terms of the weighted F1 score in the IHSC dataset, however, our method has a better performance in the positive class (Hate), which is more important in this case since the task is binary classification (Normal vs. Hate) in this dataset.

Ensemble methods based on different user styles can be useful for reflecting user styles for hate speech. We demonstrate this idea with the influential feature of tweet authors. However, the Influential method is not the highest-performing method in any datasets, probably because the users in the same ensemble member do not necessarily have the same writing style of hate speech, or due to possible noise in the ensemble members. Nevertheless, we argue that this method is still promising such that its combination version performs in the top-3 highest methods in Founta18 and Toraman22-EN. We argue that a deeper examination of data splits or user features might improve performance by reflecting different hate styles.

When we compare our normal and combination approach for constructing ensembles in Fig. 4, we observe that the Combination version outperforms in 12 of 14 (86%) comparisons. The better performance of the Combination method is probably due to the increased amount of samples in data splits used in training each ensemble member. The efficacy of the suggested models depends critically on the size of the training set while using deep learning methods (Bailly et al. Reference Bailly, Blanc, Francis, Guillotin, Jamal, Wakim and Roy2022). Note that our normal method uses a single data split (20% of the whole data) to train an ensemble member, while the Combination method merges four splits (80% of the whole data) that result in a similar number of instances as in the baseline methods (Bagging and Fold). We argue that our main ensemble approach splits the data to obtain different aspects of hate speech for ensemble learning, and the combination of different aspects can further improve the performance of hate speech detection.

Figure 4. Performance comparison of our proposed ensemble methods (i.e. GMM, Topic, and Influential) with their Combination versions, as explained in Section 3, for all datasets in different subplots.

In terms of class labels, we observe that the content-based models (i.e. GMM and Topic) perform the highest for the Hate class in four of five datasets. We argue that different writing styles can be captured by the content-based models since they cluster sentence embeddings, which are encoding vectors of the text contents.

Furthermore, we observe that the Topic method performs the highest for the Offensive/ Abusive class in three of four datasets and for the Spam class in the Founta18 dataset. Some words are not necessarily offensive, abusive, or spam in another context. We thereby argue that the topic splits can better capture the differences in offensive, abusive, and spam classes.

RQ-3: Language-independent ensemble methods for hate speech detection. The good performance of ensemble methods is also valid for other languages, that is Italian and Turkish. In our experiments, the highest-performing ensemble method improves the performance of the single approach statistically significantly in both Italian and Turkish languages. Moreover, an ensemble method that leverages different topic aspects of hate speech, GMM Combination, has the highest performance in the Hate class in both Italian and Turkish. We argue that our observations support that the performance of hate speech ensembles is language-independent, though the experiments are still limited to three languages (English, Italian, and Turkish).