The amount and impact of hate speech on social media is already alarming, causing harm for individuals and societies, yet it still keeps increasing. It has been reported that hate speech is strongly related to hate crimes and suicide rates. To stop hate speech on social media, online automatic approaches for hate speech detection in text have thus been drawing wide attention in the last decade. In this work, we take up the HASOC-21 shared tasks, which focus on hate speech and ofensive content identification in English. Two subtasks are addressed-one to identify the presence of any form of hate, ofensive and profane content in a post, the other to discriminate between these three types of problematic content. We propose a hybrid representation fusion (HRF) structure, using both TF-IDF and BERTweet-based representations. Specifically, a Convolutional-BiLSTM (CBLSTM) network is used to extract semantic information from the BERTweet embeddings. Finally, we use the cosine similarity between diferent types of features as a reliability measure. We compare the performance between the proposed model with and without cosine similarity as an additional reliability measure. The former could reduce the standard deviation of the macro f1 score in a 10-fold cross-validation by over 27.27% for subtask A and 22.22% for subtask B, respectively. In the HASOC-21 ranking, our approach achieved an 80.13% macro F1 score for subtask A and 64.82% for subtask B on the oficial HASOC-21 test set.
In recent years, digital services like social media platforms or instant messaging systems have been struggling with hateful comments and abusive content, where users—typically anonymous—justify or instigate violence and discrimination against a person or a group. Hate speech permeates the vast majority of social media platforms and instant messaging systems, and has been increasing in volume and severity [1].
Nowadays, the impact of online hate speech (and misinformation) has entered all protagonists’ consciousness, as hate speech strongly afects violent crimes in the real world. Research studies show a strong connection between the increased amount of shared and distributed hate speech messages and a higher occurrence of hate-related ofenses against minorities [ 2]. And even nEvelop-O LGOBE (D. Kolossa)
https://cognitive-signal-processing.de/index.php/team/ (W. Yu); https://cognitive-signal-processing.de/index.php/team/ (B. Boenninghof); https://cognitive-signal-processing.de/index.php/team/ (D. Kolossa) training test
NOT though not all cases of criminal acts directly follow from abusive online messages, its impact on, e.g., individual votes during democratic elections is vast [3, 4]. Additionally, victims are often strongly afected by the floods of hateful comments, and sufer from feelings of hopelessness or isolation, eating or sleep disorders, depression, and thoughts of suicide[5, 6, 7].
Consequently, there is an urgent need to find suitable tools and instruments to at least reduce the prevalence and impact of hate speech. However, due to the rising amount of shared and distributed poisonous content on social media platforms or instant messaging systems, it is infeasible to identify all, or even a significant part, manually.
The HASOC-21 shared task [8] encourages the participants working on automated solutions to counter this negative trend. The shared task ofers annotated messages to develop machine learning methods for identifying hate speech and diferentiating between diferent types of hate speech. More precisely, the dataset contains Twitter posts for both, English and Indo-Aryan languages [9]. In this work, we decided to work only with the annotated English tweets.
The paper is organized as follows: The dataset and our feature extraction are described in Section 2. Section 3 gives an overview of the proposed models, followed by the experimental setup, which is presented in Section 4. Finally, the results are discussed in Section 5 and conclusion are drawn in Section 6.
The HASOC-21 dataset consists of annotated Twitter posts, with training and test set comprising 3843 and 1281 tweets, respectively. Each tweet has been sorted into the binary categories NOT (Non-Hate-Ofensive) or HOF (Hate-Ofensive). The HOF tweets are further discriminated into the three classes HATE (Hate speech), OFFN (Ofensive), or PRFN (Profane). Table 1 shows the class distribution of the dataset. 2.1. Feature Extraction In this work, we employ two types of features, the statistical term frequency-inverse document frequency (TF-IDF) features, and semantic text embeddings extracted from the pre-trained BERTweet model [10].
Before extracting the TF-IDF features, we perform text normalization on the raw tweets to remove redundant and noisy information. Concretely, we decided to conduct the following steps: 1. Replacing all user names and hashtags with USER and HASH. 2. Removing all URLs.
3. Converting all Emojis into corresponding text representations1.
To see the efect of this normalization, consider this raw tweet: @maiysha Death threats to Actor siddharth and his family for exposing BJP What can we expect from the party which files FIR for asking Oxygen #IStandWithSiddharth https://t.co/4Fwpw4y8Bf #ModiKaVaccineJumla.
After text normalization, we obtain:
USER Death threats to Actor siddharth and his family for exposing BJP What can we expect from the party which files FIR for asking Oxygen :pouting_face: HASH HASH.
Apart from these steps, we also tested the efect of converting all upper-case tokens into lower case. However, results on the PAN18 English dataset [11], did not show any notable improvement.
The BERTweet model also provides a built-in text normalization stage, resulting in the following text-normalized tweet: @USER Death threats to Actor siddharth and his family for exposing BJP What can we expect from the party which files FIR for asking Oxygen :pouting_face: #IStandWithSiddharth HTTPURL #ModiKaVaccineJumla .
TF-IDF features are based purely on term frequencies in collections of documents. Similar to [12], the TF-IDF features are based on n-gram models, where we perform Latent Semantic Analysis (LSA) [13] to reduce the dimension of the features. We use both, token-based 1to 3-gram and character-based 3- to 5-gram models to extract the TF-IDF features, by the implementation provided in the scikit-learn library [14]. The TF-IDF features are extracted by the T f i d f V e c t o r i z e r function, which transforms each tweet into a vector2. The minimum document frequency is set to 2. The maximum document frequency is set to 1.0 (100%), which means terms occurring in all documents are ignored. Then, LSA is applied through the truncated singular value decomposition (SVD) [15]. The truncated SVD transforms the TF-IDF features to a 500-dimensional vector, which is done separately for the token- and character-based n-gram features 3. Finally, the token- and the character-based transformed n-gram TF-IDF features are concatenated to obtain a 1000-dimensional feature vector.
In addition, we also extract semantic text embeddings from the pre-trained contextualized BERTweet language model 4. As described in [10], this model was pre-trained on a large corpus 1https://github.com/carpedm20/emoji 2https://scikit-learn.org/stable/modules/generated/sklearn.feature_extraction.text.TfidfVectorizer.html 3https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.TruncatedSVD.html 4https://github.com/VinAIResearch/BERTweet.git of English tweets.
Specifically, we chose the BERTweet-large model, where the embedding dimension for a single token is 1024, so that we obtain a sequence of 1024-dimensional token embeddings for each tweet.
In this section, we present the model architecture of our final submission.
Text Text normalization
BERTweet
TF-IDF Feed-forward t:B 1000 te:B 768
CLSTM token [CLS] and ended with [SEP], e.g. “[CLS] @dxnprivv Stop being a twat then. [SEP]”. After the BERTweet embedding, the first token [CLS] representation is always used as a special classification token for text classification tasks.
Hence, we use the representation vector of the first token (i.e. [B, 0, 1024]) as BERTweet classification feature. A feed-forward network is used to map into a lower dimension.
Additionally, the complete embedding sequence is fed into a Convolutional-BiLSTM (CBLSTM)
network [17] to extract semantic information. We denote the input embedding sequence for a
tweet as x = [x1, x2, ⋯ x ], where x , ∈ {1, … , } is a 1024-dimensional embedding vector of
the -th token. In the CBLSTM network, the sequence x is first fed into 4 BiLSTM layers,
=̃ FF ( ).
ℎ(
ℎ(
⋮
ℎ(
where the cell dimension of each layer is = 768 , and h(
) ⋅ ℎ(
Here, w ∈ R1 × ⋅2 represents the convolution filter and b is the corresponding bias. The window size is chosen as ∈
{3, 4, 5}. The convolution is performed on each possible window of the
matrix h(
v = max{c1 , c2 , ⋯ , c−+1 }.
where W
given by:
After concatenating the pooling results, v = [v3, v4, v5], the attention weights are obtained by
= softmax (ℎ̃ ⋅ W ⋅ v) ,
∈ R3 × 1 is the trainable parameter vector. The final hidden state representation is
=1
s = ∑ ⋅ h(
=̃ FF ([r , ]̃). (10) (11) (12) (13)
Now, the hidden state representation is concatenated with the last state of the last hidden layer,
h(
Subsequently, the BERTweet classification vector ̃ in Eq (
Now, we can use the TF-IDF- and the BERTweet-based representations to make the decision. If the representations for these two feature types are similar, the TF-IDF model and the BERT model are tending towards the same outcome and vice versa. To help the model towards unified decisions, we calculate the cosine similarity of these two representations. Finally, the two feature representations and the cosine similarity are concatenated, and a feed-forward classifier makes the final decision based on this vector.
We evaluated our proposed model on two subtasks. subtask A is a binary task, to identify each tweet as ’NOT’ (non-ofensive) or as ’HOF’ (hate/ofensive speech). For this task, the network output dimension is thus 1 and we used a sigmoid function as the output layer activation function. The network was trained with the binary cross-entropy as the loss function. subtask B is to classify tweets into four diferent categories, namely ’HATE’ (Hate speech), ’OFFN’ (Ofensive), ’PRFN’ (Profane) or ’None’ (No hate speech). The output dimension was thus set to 4, with a softmax function as the output layer activation function and the cross-entropy loss. Each feed-forward network contains two hidden layers of 768 units each, with ReLU activation functions.
To compare the model performance in diferent setups, we applied 10-fold cross-validation on the training set. Five diferent model versions are implemented: • TF-IDF : the TF-IDF stream in the HRF model (Figure 1) only uses TF-IDF as features. • BERT-C: the BERTweet classification stream in the HRF model, based solely on the BERTweet classification features .̃ ding features r .
• BERT-E: the BERTweet embedding stream in HRF model, based on the BERTweet embed• HRF : our proposed model combining both feature streams as shown in Figure 1. • HRF-NC: our proposed model, without the cosine similarity as a reliability measure.
All models were implemented using the PyTorch library [18]. We trained the models for 50 epochs, choosing a batch size of 48, and using the AdamW optimizer [19]. To avoid overfitting, we incorporated early stopping to monitor the degradation on the held-out set. More precisely, the training was stopped when the accuracy on the validation set did not increase over 9 epochs. We set the initial learning rate to 2 ⋅ 10−5 and we further installed a linear learning rate scheduler with warm-up to change the learning rate during the first four epochs of the training stage.
During training, all trainable parameters of the pre-trained BERTweet models were also ifne-tuned. Inspired by [ 20], we finally averaged the weights of those two epoch-wise models, which returned the highest validation accuracy during the training stage.
All models were trained using NVIDIA’s Volta-based DGX-1 multi-GPU system, using 2 Tesla V100 GPUs with 32 GB memory each.
The two box plots in Figure 2 provide the cross-validation results of the baseline models with our proposed hybrid model on the training set. The validation set is used for early stopping. As it can be seen, the F1 score of the TF-IDF -based model that only relies on traditional features is significantly lower than that of the other models. We use the Mann-Whitney-U test [ 21] to indicate statistically significant diferences in the F1 score between the proposed HRF model and other baselines. However, from Plots 2a and 2b, we do not notice an improvement from the combination of both, traditional and neural features.
Table 2 shows all obtained F1 scores in greater detail. Starting with subtask A, the BERT -based models again significantly outperform the TF-IDF system, while BERT-C is on par with HRF. For subtask B, the proposed hybrid model slightly outperforms the BERT-C model. However, overall, for both subtask A and subtask B, there is no statistically significant diference between the diferent BERT -based models.
After getting access to the oficial groundtruth labels, we repeated the cross-validation procedure on the HASOC-21 test set. The results are shown in Figure 3 and Table 3. Altogether, we do not observe significant diferences between the training and the test sets, which shows the robustness of our approaches. Having a closer look at Table 3, the HRF model reduces the standard deviation of the macro F1 score by over 27.27% for subtask A and 22.22% for subtask B compared to the HRF-NC model. However, on subtask B, the BERT-E model has the better performance. Figure 3a still shows no statistically significant diference between the BERT -based models, while we can see a slight diference between BERT-C and the proposed HRF model in Figure 3b.
Finally we submitted diferent versions of the HRF model to participate in the HASOC-21 ranking. We randomly selected 128 Twitter posts from the training set as the validation set. The model performance is tested on the HASOC-21 oficial test set.
For the final raking, all participants were allowed a maximum of five submissions. Firstly, we submitted two diferent versions of the HRF system, with and without fine-tuned BERTweet parameters. The first two rows in Table 4 show these results for subtask A and emphasise the advantage of task-specific fine-tuning of the standard BERTweet model. Next, we submitted diferent HRF versions (for both tasks) which are based on the standard or large BERTweet version. It is observable that the large BERTweet model tends towards a slightly better performance, especially for subtask B.
In summary, our best models in Table 4 achieved rank 11 out of 56 participants for subtask A and placed 3rd out of 37 participants for subtask B in the HASOC-21 ranking [8].
We are presenting our approach for the HASOC-21 shared tasks of hate speech identification and classification, detailing its stages from the pre-processing steps, via the design of the neural model architecture, to the training procedure. The core of our approach is a novel neural fusion strategy to combine classical TF-IDF features with fine-tuned contextualized BERTweet embeddings. We incorporated the cosine similarity between the final TF-IDF- and BERT-based representations to learn a suitable fusion network. The results of our 10-fold cross-validation are consistent with the test set results. In the challenge, our submissions ranked 11th out of 56 participants for subtask A and 3rd out of 37 participants for subtask B.
The vast amounts of information on social media cannot realistically be monitored for hate speech by linguists. To address the demand of this situation, we would like to design a model that focuses not only on accuracy but also on delivering reliable confidence values. In this context, and moving one step towards responsible AI, future work will focus on calibration methods such as Dirichlet calibration [22]. These are important for obtaining reliable output predictions, in which confidence is well-matched with accuracy, so that the system output provides a sound basis for real-world decision-making.
The work was supported by the PhD School ”SecHuman - Security for Humans in Cyberspace” by the federal state of NRW, and partially funded by the Deutsche Forschungsgemeinschaft (DFG – German Research Foundation) - Project-ID 429873205. [7] N. Kteily, E. Bruneau, Backlash: The politics and real-world consequences of minority group dehumanization, Personality and Social Psychology Bulletin (2017). [8] S. Modha, T. Mandl, G. K. Shahi, H. Madhu, S. Satapara, T. Ranasinghe, M. Zampieri, Overview of the HASOC Subtrack at FIRE 2021: Hate Speech and Ofensive Content Identification in English and Indo-Aryan Languages and Conversational Hate Speech, in: FIRE 2021: Forum for Information Retrieval Evaluation, Virtual Event, 13th-17th December 2021, ACM, 2021. [9] T. Mandl, S. Modha, G. K. Shahi, H. Madhu, S. Satapara, P. Majumder, J. Schäfer, T. Ranasinghe, M. Zampieri, D. Nandini, A. Jaiswal, Overview of the HASOC subtrack at FIRE 2021: Hate Speech and Ofensive Content Identification in English and Indo-Aryan Languages, in: Working Notes of FIRE 2021 - Forum for Information Retrieval Evaluation, CEUR, 2021.
URL: http://ceur-ws.org/. [10] D. Nguyen, T. Vu, A. Nguyen, BERTweet: A pre-trained language model for English
Tweets, 2020. [11] F. Rangel, P. Rosso, M. M. Gómez, M. Potthast, B. Stein, Overview of the 6th author profiling task at PAN 2018: Multimodal gender identification in Twitter, Working Notes Papers of the CLEF (2018) 1–38. [12] S. Daneshvar, D. Inkpen, Gender identification in twitter using n-grams and LSA, in: Proc.
CLEF 2018), 2018. [13] S. T. Dumais, Latent semantic analysis, Annual review of information science and technology (2004). [14] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al., Scikit-learn: Machine learning in Python, the Journal of machine Learning research 12 (2011). [15] N. Halko, P. Martinsson, J. Tropp, Finding structure with randomness: Stochastic algorithms for constructing approximate matrix decompositions (2009). [16] J. Devlin, M. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, arXiv preprint arXiv:1810.04805 (2018). [17] J. Zheng, L. Zheng, A hybrid bidirectional recurrent convolutional neural network attention-based model for text classification (2019). [18] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al., Pytorch: An imperative style, high-performance deep learning library, Advances in neural information processing systems (2019). [19] I. Loshchilov, F. Hutter, Decoupled weight decay regularization, arXiv preprint arXiv:1711.05101 (2017). [20] G. Huang, Y. Li, G. Pleiss, Z. Liu, J. E. Hopcroft, K. Q. Weinberger, Snapshot ensembles: Train 1, get M for free, in: 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings, 2017. [21] H. Mann, D. Whitney, On a test of whether one of two random variables is stochastically larger than the other, The annals of mathematical statistics (1947) 50–60. [22] M. Kull, M. Perello-Nieto, M. Kängsepp, H. Song, P. Flach, et al., Beyond temperature scaling: Obtaining well-calibrated multiclass probabilities with dirichlet calibration, arXiv preprint arXiv:1910.12656 (2019).