=Paper=
{{Paper
|id=Vol-3159/T1-40
|storemode=property
|title=Combining Textual Features for the Detection of Hateful and Offensive Language
|pdfUrl=https://ceur-ws.org/Vol-3159/T1-40.pdf
|volume=Vol-3159
|authors=Sherzod Hakimov,Ralph Ewerth
|dblpUrl=https://dblp.org/rec/conf/fire/HakimovE21
}}
==Combining Textual Features for the Detection of Hateful and Offensive Language==
https://ceur-ws.org/Vol-3159/T1-40.pdf
Combining Textual Features for the Detection of
Hateful and Offensive Language
Sherzod Hakimov1,2 , Ralph Ewerth1,2
1
TIB - Leibniz Information Centre for Science and Technology, Hannover, Germany
2
Leibniz University Hannover, L3S Research Center, Hannover, Germany
Abstract
The detection of offensive, 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 different textual features for the detection of hateful or offensive 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, offensive and profane content from the post datasets of HASOC-2021 dataset under
the team name TIB-VA. We compared different variants of the contextual word embeddings combined
with the character level embeddings and the encoding of collected hate terms.
Keywords
hate speech detection, offensive language detection, abusive language detection, social media mining
1. Introduction
The detection of hateful, offensive and profane language has become a significant research
challenge with the widespread usage of social media. Certain groups of people become targets
of cyberbullying activities on a daily basis on many social networks such as Facebook, Twitter,
or Instagram [1]. There have been many efforts by the research community and social media
companies such as Facebook1 and Twitter2 to the scope of hate speech and tackle the problem.
In general, hate speech is defined as a language used to express hatred towards a targeted group or
individuals based on specific characteristics such as religion, ethnicity, origin, sexual orientation,
gender, physical appearance, disability or disease [2, 3, 4, 5, 6, 7, 8].
In this paper, we analyze the effects of combining multiple textual features to detect hateful,
offensive or profane language expressed in the tweet text. We evaluated our approach on the
Hate Speech and Offensive Content Identification in English and Indo-Aryan Languages (HASOC)
challenge datasets3 . We submitted our solution to the English Subtask 1A: Identifying Hate,
offensive and profane content from the post [9] of the HASOC-2021 [10] challenge series. The
task involves classifying a given tweet text whether the content is hateful, offensive, or profane
Forum for Information Retrieval Evaluation, December 13-17, 2021, India
" sherzod.hakimov@tib.eu (S. Hakimov); ralph.ewerth@tib.eu (R. Ewerth)
0000-0002-7421-6213 (S. Hakimov); 0000-0003-0918-6297 (R. Ewerth)
Β© 2021 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings CEUR Workshop Proceedings (CEUR-WS.org)
http://ceur-ws.org
ISSN 1613-0073
1
https://www.facebook.com/communitystandards/hate_speech
2
https://help.twitter.com/en/rules-and-policies/hateful-conduct-policy
3
https://hasocfire.github.io
language or not. We proposed a combination of multiple textual features based on neural
network architecture and evaluated different 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 eval-
uations of model architectures are described in detail. Finally, the Section 4 concludes the
paper.
2. Approach
Our model architecture is built on top of three textual features that are combined to predict
whether a given text contains hateful, offensive 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 different 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 different 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.
3. Experimental Setup and Results
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.
3.1. Datasets
Our model architecture is built for the HASOC-2021 [10] English Subtask 1A: Identifying Hate,
offensive and profane content from the post [9]. In Table 1, we provide the number of data points
for HASOC-2019[11], HASOC-2020 [12], and HASOC-2021 [10] editions of the challenge. The
4
https://github.com/sherzod-hakimov/HASOC-2021---Hate-Speech-Detection
5
https://www.noswearing.com/dictionary & https://hatebase.org/
BERT
RNN
Encoder
token1
token2
Character Softmax
token3 RNN
Encoder
.
.
Block-2 Block-3 Block-4
.
tokenn Hate and Offensive
Hate Words
Encoder Concatenation Not Hate or Offensive
Block-1
1D Features Linear Layer Batch ReLU
Normalization
Figure 1: The model architecture combines character, hate words, and BERT embeddings that outputs
probability of a given text being hate and offensive or not.
datasets include tweet text as input data and two target labels: Hate and Offensive (HOF) and
Not Hate or Offensive (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.
Table 1
Distribution of data points for train and test splits for English Subtask 1A for all editions of the HASOC
datasets. HOF: Hate and Offensive, NOT: Not Hate or Offensive
Train Test
Dataset
HOF NOT HOF NOT
HASOC-2019 [11] 2261 3591 288 865
HASOC-2020 [12] 1856 1852 807 785
HASOC-2021 [10] 2501 1342 765 456
3.2. Data Preprocessing
There are several challenges with working text from Twitter. In many cases, the tokens are
written in different 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.
3.3. Model Parameters
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 different 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 different types of RNN
layers: Long-short Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional Gated
Recurrent Unit (Bi-GRU). We also experimented with different 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 .
3.4. Results
We tested different model configurations as explained above for all three datasets. The results
are given in Table 2. The official evaluation metrics for the English Subtask 1A [9] is Macro
F1-score. Additionally, we included accuracy and weighted F1-score since the number of data
points for each class are not balanced for test splits of the datasets (see Table 1). We included the
best performing models with the corresponding features. Based on the initial experiments, the
choice of Gated Recurrent Unit (GRU) with the layer size of 100 yielded the highest performance
on the validation set for three datasets in comparison to other model configurations. Therefore,
all model configurations listed in the table below use the GRU layer with the size 100.
The results suggest that BERT-base embeddings have greater a impact than HateBERT embed-
dings. Another important observation is that the feature based on the multi-hot encoding of hate
terms (π»π ) achieves high accuracy and weighted F1-score for all datasets. Specifically, every
model configuration that included the π»π 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 [9] of the HASOC-
2021 [10] dataset. We submitted the same model as a team TIB-VA to the official challenge. Our
model was ranked at the position 33 with the Macro-F1 score of 0.76.
6
https://github.com/sherzod-hakimov/HASOC-2021---Hate-Speech-Detection
Table 2
Evaluation results of various model configurations on English Subtask 1A of three datasets. The eval-
uation metrics are accuracy (Acc), Macro F1-score (M-F1), and weighted F1-score (W-F1). The best
performing model configurations for each dataset are highlighted in bold. BB: word embeddings ex-
tracted from a pre-trained BERT-base [13] model, HB: word embeddings extracted from a pre-trained
HateBERT [15] model, CH: character level embeddings, HW: multi-hot encoding of hate words.
HASOC-2019 HASOC-2020 HASOC-2021
Features Acc M-F1 W-F1 Acc M-F1 W-F1 Acc M-F1 W-F1
π΅π΅ 0.78 0.70 0.78 0.86 0.86 0.86 0.74 0.73 0.73
π»π΅ 0.64 0.61 0.61 0.84 0.84 0.84 0.74 0.73 0.74
πΆπ» 0.36 0.36 0.36 0.56 0.51 0.60 0.63 0.59 0.64
π»π 0.75 0.61 0.77 0.89 0.89 0.89 0.71 0.71 0.71
πΆπ» + π»π 0.76 0.59 0.80 0.89 0.89 0.89 0.71 0.71 0.71
π΅π΅ + π»π 0.78 0.68 0.80 0.88 0.88 0.88 0.78 0.76 0.78
π΅π΅ + πΆπ» 0.77 0.70 0.77 0.81 0.81 0.81 0.73 0.72 0.73
π»π΅ + π»π 0.71 0.65 0.70 0.89 0.89 0.89 0.78 0.77 0.78
π»π΅ + πΆπ» 0.58 0.56 0.55 0.87 0.87 0.87 0.73 0.71 0.73
π΅π΅ + πΆπ» + π»π 0.75 0.69 0.75 0.89 0.89 0.89 0.79 0.77 0.79
π»π΅ + πΆπ» + π»π 0.75 0.69 0.75 0.49 0.33 0.66 0.78 0.76 0.78
We present the confusion matrices in Figure 2 for two models with different 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: π΅π΅ + πΆπ» + π»π (b) Model with features: π»π΅ + πΆπ» + π»π
Figure 2: Confusion matrices for two models evaluated on the HASOC 2021 - English Subtask 1A
4. Conclusion
In this paper, we have analyzed model architectures that combine multiple textual features to
detect hateful, offensive 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).
References
[1] R. M. Kowalski, G. W. Giumetti, A. N. Schroeder, M. R. Lattanner, Bullying in the digital age:
A critical review and meta-analysis of cyberbullying research among youth., Psychological
bulletin 140 (2014) 1073.
[2] P. Badjatiya, S. Gupta, M. Gupta, V. Varma, Deep learning for hate speech detection in
tweets, in: R. Barrett, R. Cummings, E. Agichtein, E. Gabrilovich (Eds.), Proceedings of the
26th International Conference on World Wide Web Companion, Perth, Australia, April
3-7, 2017, ACM, 2017, pp. 759β760.
[3] T. Davidson, D. Warmsley, M. W. Macy, I. Weber, Automated hate speech detection and the
problem of offensive language, in: Proceedings of the Eleventh International Conference
on Web and Social Media, ICWSM 2017, MontrΓ©al, QuΓ©bec, Canada, May 15-18, 2017,
AAAI Press, 2017, pp. 512β515.
[4] P. Fortuna, S. Nunes, A survey on automatic detection of hate speech in text, ACM Comput.
Surv. 51 (2018) 85:1β85:30. doi:10.1145/3232676.
[5] W. Warner, J. Hirschberg, Detecting hate speech on the world wide web, in: Proceedings
of the second workshop on language in social media, 2012, pp. 19β26.
[6] F. Poletto, V. Basile, M. Sanguinetti, C. Bosco, V. Patti, Resources and benchmark corpora
for hate speech detection: a systematic review, Lang. Resour. Evaluation 55 (2021) 477β523.
URL: https://doi.org/10.1007/s10579-020-09502-8. doi:10.1007/s10579-020-09502-8.
[7] Z. Zhang, L. Luo, Hate speech detection: A solved problem? the challenging case of long
tail on twitter, Semantic Web 10 (2019) 925β945. URL: https://doi.org/10.3233/SW-180338.
doi:10.3233/SW-180338.
[8] P. Zeinert, N. Inie, L. Derczynski, Annotating online misogyny, in: C. Zong, F. Xia,
W. Li, R. Navigli (Eds.), Proceedings of the 59th Annual Meeting of the Association
for Computational Linguistics and the 11th International Joint Conference on Natural
Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, Au-
gust 1-6, 2021, Association for Computational Linguistics, 2021, pp. 3181β3197. URL:
https://doi.org/10.18653/v1/2021.acl-long.247. doi:10.18653/v1/2021.acl-long.247.
[9] T. Mandl, S. Modha, G. K. Shahi, H. Madhu, S. Satapara, P. Majumder, J. SchΓ€fer, T. Ranas-
inghe, M. Zampieri, D. Nandini, A. K. Jaiswal, Overview of the HASOC subtrack at FIRE
2021: Hate Speech and Offensive Content Identification in English and Indo-Aryan Lan-
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 Offensive 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 offensive content identification in indo-
european languages, in: P. Majumder, M. Mitra, S. Gangopadhyay, P. Mehta (Eds.), FIRE β19:
Forum for Information Retrieval Evaluation, Kolkata, India, December, 2019, ACM, 2019, pp.
14β17. 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 offensive 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.
29β32. 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 trans-
formers 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 Computa-
tional 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. 4171β4186.
[14] R. Gomez, J. Gibert, L. GΓ³mez, 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. 1459β1467. 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.