The detection of ofensive, hateful and profane language has become a critical challenge since many users in social networks are exposed to cyberbullying activities on a daily basis. In this paper, we present an analysis of combining diferent textual features for the detection of hateful or ofensive posts on Twitter. We provide a detailed experimental evaluation to understand the impact of each building block in a neural network architecture. The proposed architecture is evaluated on the English Subtask 1A: Identifying Hate, ofensive and profane content from the post datasets of HASOC-2021 dataset under the team name TIB-VA. We compared diferent variants of the contextual word embeddings combined with the character level embeddings and the encoding of collected hate terms.
language or not. We proposed a combination of multiple textual features based on neural network architecture and evaluated diferent configurations. Our experimental evaluation is performed on all three datasets: HASOC-2019 [11], HASOC-2020 [12], HASOC-2021 [10].
The remainder of the paper is structured as follows. In Section 2, we describe the proposed model architecture. In Section 3, the experimental setup, challenge datasets, as well as evaluations of model architectures are described in detail. Finally, the Section 4 concludes the paper.
Our model architecture is built on top of three textual features that are combined to predict whether a given text contains hateful, ofensive or profane language. The neural network architecture is shown in Figure 1. Input tokens are fed into BERT, Character and Hate Words encoders to extract feature-specific vector representations. Once each feature representation is extracted, the outputs are fed into separate components to obtain one-dimensional vector representations. These vectors are concatenated and fed into three diferent blocks to obtain binary class probabilities. Each block is composed of a linear layer, batch normalization and a ReLU activation function. The source code and the resources described below are shared publicly with the community4. Next, we describe the textual encoders in detail. BERT Encoder: We used a pre-trained BERT [13] model to obtain contextual 768-dimensional word vectors for each input token.
Character Encoder: Each input token is converted into vector representation based on the one-hot encoding of characters in English. We only use letters (a-z) to obtain a sequence of character-level vectors.
Hate Words Encoder: We collected a list of hate terms by combining the dictionary provided by Gomez et al. [14] with additional online dictionaries5. We manually filtered out terms that do not express hate concepts and obtained a list of 1493 hate terms. The list contains a variety of terms with diferent lexical variations to increase the coverage of detecting such terms in tweets, e.g., bitc*, border jumper, nig**, or chin*. This encoder outputs a 1493-dimensional vector, a multi-hot encoding of hate terms in input tokens.
In this section, we present the challenge datasets, details about data preprocessing and model parameters, and finally explain the experimental setup along with the obtained results.
Our model architecture is built for the HASOC-2021 [10] English Subtask 1A: Identifying Hate,
ofensive and profane content from the post [
... tokenn Character
Encoder Hate Words
Encoder RNN
Block-1
Block-2
Block-3 Block-4 Concatenation
Softmax Hate and Offensive Not Hate or Offensive 1D Features
Linear Layer
Batch Normalization
ReLU datasets include tweet text as input data and two target labels: Hate and Ofensive (HOF) and Not Hate or Ofensive (NOT) .
The number of training samples for the 2019 and 2021 editions are not equally distributed among the two classes. To overcome the class imbalance issue, we applied the oversampling method to the training splits. We randomly selected a certain number of data points for the minority class (HOF for 2019, NOT for 2021) and duplicated them to equalize with the number of data points for the majority class.
There are several challenges with working text from Twitter. In many cases, the tokens are written in diferent forms to save space, capitalized, mixed with numbers etc. We apply the following text preprocessing steps to normalize the tweet text: 1) remove hashtags, URLs, user tags, retweets tags using Ekphrasis, 2) remove punctuations, 3) convert tokens into lowercase.
In this section, we provide details about all parameters of the buildings blocks in the model architecture show in Figure 1.
BERT Encoder: We experiment with two diferent variants of BERT models. The first variant is BERT-base, which is the default model provided by Devlin et al. [13]. The second variant is HateBERT provided by Caselli et al. [15], which is a BERT-base model pre-trained further on hateful comments corpus extracted from Reddit. Both variants output a sequence of 768-dimensional vectors for the given input tokens.
Recurrent Neural Network (RNN) Layers: We experimented with diferent types of RNN layers: Long-short Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional Gated Recurrent Unit (Bi-GRU). We also experimented with diferent layer sizes, which are 100, 200, 300.
Linear Layers: The model architecture includes four blocks that are composed of three consecutive layers: linear layer, batch normalization, and activation function (ReLU). The sizes of the linear layers in the Block-1, Block-2, Block-3, Block-4 are 512, 512, 256, 128 respectively.
Training Process: Each configuration of the model architecture is trained using Adam optimizer [16] with a learning rate of 0.001, a batch size of 64 for maximum of 20 iterations. We use the 90:10 splits for train and validation splits to find the optimal hyperparameters.
Implementation: The model architecture is implemented in Python using the Tensorflow Keras library. The source code is shared openly with the community6.
We tested diferent model configurations as explained above for all three datasets. The results
are given in Table 2. The oficial evaluation metrics for the English Subtask 1A [
The results suggest that BERT-base embeddings have greater a impact than HateBERT
embeddings. Another important observation is that the feature based on the multi-hot encoding of hate
terms ( ) achieves high accuracy and weighted F1-score for all datasets. Specifically, every
model configuration that included the feature yields the best results on the HASOC-2020
dataset. Our approach that combines BERT-base, character embeddings, and multi-hot encoding
of hate terms achieved a Macro F1-score of 0.77 on English Subtask 1A [
6https://github.com/sherzod-hakimov/HASOC-2021---Hate-Speech-Detection
We present the confusion matrices in Figure 2 for two models with diferent pre-trained variants of BERT models: BERT-base (BB) and HateBERT (HB). Both models were trained with character level embeddings (CH) and multi-hot encoded hate words (HW). We can observe that the model using BERT-base embeddings (Figure 2a) makes more correct predictions (614 vs. 577) in detecting hateful content (HOF) when compared with the other model variant (Figure 2b). A similar pattern exists for cases where the target class is NOT, and the model predicts HOF where the model with the BB feature makes fewer mistakes (151 vs. 188) than the other model. (a) Model with features: + + (b) Model with features: + +
In this paper, we have analyzed model architectures that combine multiple textual features to detect hateful, ofensive and profane language. Our experimental results showed that simply using the multi-hot encoding of collected 1493 hate terms yields significant performance. The combination of BERT embeddings, character embeddings, and features based on hate terms achieved the best performance on the English Subtask 1A, HASOC 2021 dataset. Another observation of the evaluation is that a variant of the BERT model trained on domain-specific (HateBERT ) text did not improve the results in comparison to the default pre-trained model variant (BERT-base). guages, in: Working Notes of FIRE 2021 - Forum for Information Retrieval Evaluation, CEUR, 2021. URL: http://ceur-ws.org/. [10] 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. [11] T. Mandl, S. Modha, P. Majumder, D. Patel, M. Dave, C. Mandalia, A. Patel, Overview of the HASOC track at FIRE 2019: Hate speech and ofensive content identification in indoeuropean languages, in: P. Majumder, M. Mitra, S. Gangopadhyay, P. Mehta (Eds.), FIRE ’19: Forum for Information Retrieval Evaluation, Kolkata, India, December, 2019, ACM, 2019, pp. 14–17. URL: https://doi.org/10.1145/3368567.3368584. doi:10.1145/3368567.3368584. [12] T. Mandl, S. Modha, A. K. 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: P. Majumder, M. Mitra, S. Gangopadhyay, P. Mehta (Eds.), FIRE 2020: Forum for Information Retrieval Evaluation, Hyderabad, India, December 16-20, 2020, ACM, 2020, pp. 29–32. URL: https://doi.org/10.1145/3441501.3441517. doi:10.1145/3441501.3441517. [13] J. Devlin, M. Chang, K. Lee, K. Toutanova, BERT: pre-training of deep bidirectional transformers for language understanding, in: J. Burstein, C. Doran, T. Solorio (Eds.), Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), Association for Computational Linguistics, 2019, pp. 4171–4186. [14] R. Gomez, J. Gibert, L. Gómez, D. Karatzas, Exploring hate speech detection in multimodal publications, in: IEEE Winter Conference on Applications of Computer Vision, WACV 2020, Snowmass Village, CO, USA, March 1-5, 2020, IEEE, 2020, pp. 1459–1467. URL: https: //doi.org/10.1109/WACV45572.2020.9093414. doi:10.1109/WACV45572.2020.9093414. [15] T. Caselli, V. Basile, J. Mitrovic, M. Granitzer, Hatebert: Retraining BERT for abusive language detection in english, CoRR abs/2010.12472 (2020). URL: https://arxiv.org/abs/ 2010.12472. arXiv:2010.12472. [16] D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, in: Y. Bengio, Y. LeCun (Eds.), 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. URL: http://arxiv.org/abs/ 1412.6980.