=Paper=
{{Paper
|id=Vol-3159/T1-14
|storemode=property
|title=Classification of Hate Speech and Offensive Content using an approach based on DistilBERT
|pdfUrl=https://ceur-ws.org/Vol-3159/T1-14.pdf
|volume=Vol-3159
|authors=Swetha Saseendran,Sudharshan R,Sreedhar V,Sharan Giri
|dblpUrl=https://dblp.org/rec/conf/fire/SaseendranRVG21
}}
==Classification of Hate Speech and Offensive Content using an approach based on DistilBERT==
https://ceur-ws.org/Vol-3159/T1-14.pdf
Classification of Hate Speech and Offensive Content
using an approach based on DistilBERT
Swetha Saseendran, Sudharshan R., Sreedhar V. and Sharan Giri
Sri Sivasubramaniya Nadar College of Engineering, Chennai, Tamil Nadu, India
Abstract
This paper describes the research that our team,’Binary Beings’, did on the shared task HASOC, con-
ducted by FIRE-2021, which involves identification of hate and offensive language in various comments
on social media. Our task is divided into two hierarchical sub-tasks on an English dataset. We employed
and compared various Machine Learning, Deep Learning techniques and used pre-trained models to un-
derstand which model is most accurate in predicting the classes. Our best model [DistilBERT] obtained
a Macro F1 score of 74.91% for SubTask-A and 57.65% for SubTask-B..
Keywords
Machine Learning, LSTM, DistilBERT, Cross-Validation, TF-IDF, Multi-class Classification
1. Introduction
Social media today is a hotbed of curbing hate speech. This has emerged as a critical challenge
for governments globally0 . People are spending a considerable amount of time on social media
like Facebook, Twitter, Instagram. Studies suggest that most of the online content generated
on these platforms contain abusive language. There is a need to develop adequate response
mechanisms in order to find a balance between freedom of expression on one side, the ability
to live without oppressive remarks on the other and a requirement for a robust technology to
identify problematic content automatically.
HASOC provides a forum for developing and testing text classification systems for various
languages. It organized a shared task for FIRE 2021. The task is aimed at identifying hateful and
offensive language in social media posts. The task was organized for two languages namely
English and Hindi, but we conducted our investigation only on the English dataset. [8]
Forum for Information Retrieval Evaluation, December 13-17, 2021, India
" swetha18183@cse.ssn.edu.in ( Swetha Saseendran); sudharshan18173@cse.ssn.edu.in ( Sudharshan R.);
sreedhar18161@cse.ssn.edu.in ( Sreedhar V.); sharan18141@cse.ssn.edu.in ( Sharan Giri)
~ https://github.com/swetha4444 ( Swetha Saseendran); https://github.com/exploring-curiosity ( Sudharshan R.);
https://github.com/sreedhr92 ( Sreedhar V.); https://github.com/sharan0276 ( Sharan Giri)
© 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
http://ceur-ws.org
ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org)
There were two tasks given
Task 1: Identification and discriminaton of hate, offensive and profane content from the post.
Task 2: Identification of conversational hate-speech in code-mixed languages.
Of these we chose to do Task 1. The problem statement we chose consists of hierarchical
two-subtasks:
• SUB-TASK A: Identification of hate and offensive language, posts are classified as follows:
– Hate and Offensive(HOF)
– Not Hate and Offensive(NOT)
• SUB-TASK B: Categorization of Hate and Offensive, posts are further categorized as:
– Hate Speech(HATE)
– Offensive(OFFN)
– Profane(PRFN)
Various ML and DL models have been surveyed to check which produces more accuracy
in detecting the various Hate and Offensive comments. A pre-trained DistilBERT model is
also used and it also inspired us to propose our own model with the help of DistilBERT and
other Profanity Detection libraries (Profanity filter and better-profanity) for classifying the HOF
comments detected from Sub-Task A, as HATE,OFFN or PRFN in Sub-Task B.
2. Exploratory Data Analysis
The datasets provided were annotated in a hierarchical fashion as shown below: Table 1: .Class
Distribution of the data across sub-tasks
Table 1. Details of Dataset Provided
Details Posts in Train Data Posts in Test Data
SUB-TASK A 𝑇 𝑜𝑡𝑎𝑙 = 3843
Hate and Offensive posts (HOF) 2501 1281
Non Hate and Offensive posts (NOT) 1342
SUB-TASK B 𝑇 𝑜𝑡𝑎𝑙 = 3843
Hate (HATE) 683
Offensive (OFFN) 622 1281
Profane (PRFN) 1196
NONE 1342
In the training data for Subtask-A, 65% of the data are hate comments (HOF) and about 35%
are non-hate comments (NOT). For Subtask-B, the 65% hate comments were classified as HATE
(18% of the HOF comments), OFFN (16% of the HOF comments) and PRFN (31% of the HOF
comments).
The following images are the word-cloud for various types of comments given in the dataset.
It is a graphical representation of word frequency of different words used in each category of
the comments
0 https://www.hindustantimes.com/analysis/it-is-time-to-regulate-hate-speech-on-social-
media/story-x2JfnAcZ4mh404CM2wQLpO.html
Fig 1: Word Cloud for HATE comments Fig 2: Word Cloud for OFFN comments
Fig 3: Word Cloud for PRFN comments Fig 4: Word Cloud for NONE comments
Fig 5: Meta Data Analysis of the training data
This explains the meta data analysis for training data. It summarizes basic information about
data, making finding & working with particular instances of data easier. The diagram shows
the distribution of unique words, the mean word length and character count of each type of
comment over the training data set.
3. Methodology
Fig 6: Overall flow and methodology of the experiment
The above diagram depicts the flow of the experiment from pre-processing till classifying the
text using state-of-the-art techniques.
3.1. Pre-Processing
Natural language refers to a language that has formed and evolved naturally over a long
period, such as Hindi and English, and is commonly spoken and used. Natural language
processing analyzes the meaning of natural language to enable computers to process the
language. Natural language processing is applied in areas such as text classification, sentiment
analysis, summarization, and text clustering. This processing includes three steps:
1. Text Collection
2. Text preprocessing
3. Machine learning model
In the first step (text collection), the texts to be processed are collected. The second step (text
preprocessing) involves the standardization of unstructured texts to increase the accuracy of
natural language processing. The text collected contains many elements that are difficult to
analyze, such as tags, references, abbreviations. Most are expressed as if speaking in a care-free
manner, in terms of the vocabulary or the structural order of the sentence. Therefore, after text
processing, including changing the uppercase to lowercase, deleting special characters, tags,
@’s, and removing stopwords, preprocessing is performed according to the requirement and
stemming is also done. Finally, in the machine learning modelling stage, a supervised learning
model is established, and training and prediction are performed using vectorized number-type
data using TF-IDF. In this study, we use the LSTM, BERT and other basic ML models for training
and prediction, for hate,offense and profanity detection.
3.2. Feature Extraction
3.2.1. TF-IDF
Initially, BoW was used for training. However, it was quite inefficient in training the models
owing to the fact that BoW doesn’t consider term ordering and the rareness of the term. Hence,
we used TF-IDF to overcome some of these drawbacks.The TF-IDF model contains information
on the more important words and the less important ones as well. Thus performed well on the
ML models.
3.2.2. BERT Encoding
The ktrain2 library is used which contains the Transformers API, that allows the use of any Hug-
ging Face transformers model. And we used DistilBERT. The preprocess-train and preprocess-
test functions were used for the model name ’distilbert-base-uncased’. Word embeddings are
generated from the transformer model which is used for encoding.
ktrain is a lightweight wrapper for the deep learning library TensorFlow Keras to help build,
train, and deploy neural networks and other machine learning models and to make deep learning
and AI more accessible and easier to apply.
3.3. Classifiers
3.3.1. ML Methods
For our experiment we used 3 approaches. First, a Frequency dictionary1 was built, with which
we train the model and obtain the theta parameter for the sigmoid function that would better
represent the data. The accuracy for this model came to 66%. To improve the accuracy we went
ahead with TF-IDF Vectorizer, which was used to train a logistic regression model. The model
was imported from the Scikit-learn python package. This improved the accuracy of the model
to about 80%.
1 https://towardsdatascience.com/sentiment-analysis-using-logistic-regression-and-naive-
bayes-16b806eb4c4b
For the next two models, we used a cross-validation approach with ten folds to train different
models and saved the model with the best accuracy score. This validation feature was imported
from the scikit-learn model-selection package. For the second approach, we went ahead with an
SVM model.Given the task was a classification, SVM seemed to be an easier and efficient way to
classify the tweet sentiments. The best model that was saved had an accuracy of about 80%.[5]
To enable classification beyond binary scope, we went ahead to apply the K0.5 Nearest
Neighbor Classification technique. We used the Euclidean Distance as the criterion to determine
the clusters that would be formed.This was used specifically for Task B where there were four
labels for classification namely (PRFN, OFFN, HATE, NONE). The validation accuracy of this
model came to about 75% for Sub-Task A and 64% for Sub-Task B.
The last model, we went ahead with Random Forest Classifier with entropy criterion. Random
Forest tends to behave better in the case of noisy data as well. This pushed the accuracy of our
model to about 67% with SubTask A.
Finally, we combined the predictions of all models together and took the majority of the
prediction (mode) as the final say for Sub-Task B.
Table 2: Validation accuracy results for ML Models Sub-Task A :
Model Accuracy Macro-F1 Macro-Precision Macro-Recall
Logistic Regression 80% 74% 80% 73%
SVC 80% 77% 76% 77%
KNN 75% 69% 72% 68%
Random Forest 80% 77% 76% 77%
Table 3: Validation accuracy results for ML Models Sub-Task B :
Model Accuracy Macro-F1 Macro-Precision Macro-Recall
SVC 60% 52% 52% 53%
KNN 57% 43% 49% 46%
Random Forest 58% 49% 50% 50%
3.3.2. LSTM
Long Short-Term Memory (LSTM) is a specific recurrent neural network (RNN) architecture that
was designed to model temporal sequences and their long-range dependencies more accurately
than conventional RNNs. We used LSTM based neural network classifiers for both Sub-Task
A and Sub-Task B. We used word tokens, Embedding layer (256 dimensions), input length
(2500 words) as the inputs to the LSTM (64 units) layer with dropout of 20% and recurrent
dropout of 20%, softmax layer (for prediction) in the keras toolkit. In this pipeline, we used
binary-crossentropy as loss function, the Adam optimizer to optimize the parameters.
2 https://pypi.org/project/ktrain/
Fig 7 : LSTM model for Sub-Task A Fig 8: LSTM model for Sub-Task B
Table 4: Validation accuracy results for LSTM Model:
Task Accuracy Macro-F1 Macro-Precision Maro-Recall
Subtask - A 78% 74% 76% 73%
Subtask - B 60% 17% 19% 21%
3.3.3. DistilBERT
Distilbert runs 60% faster while preserving over 95% of BERT’s performance. It also uses less
parameters than BERT. It outperformed BERT and has now cemented itself as the model to beat
for not only text classification, but also advanced NLP tasks. Implementation of DistilBERT
pipeline is much easier than using any other pre-trained learning and thus helpful in performing
transfer learning.Hence, we chose to use DistilBERT.3
It contains 6 layers, 768 dimensions and 12 heads with 66 Million parameters and is pretrained
on BookCorpus, a dataset consisting of 11,038 unpublished books and English Wikipedia3 . For
the English dataset, we used a pre-trained ‘distilbert-base-uncased’ model for both Sub-Task A
and Sub-Task B. The model was trained and validated with the validation dataset and finally
tested on the test set given. For training the max-len of the input sequence was chosen to be
500. If you set the max-length very high, you might face memory shortage problems during
execution. On testing the model on validation dataset (15% split from training dataset) the
following results were inferred:
Table 5: Validation accuracy results for DistilBERT Model:
Task Accuracy Macro-F1 Macro-Precision Maro-Recall
Subtask - A 78% 73% 75% 72%
Subtask - B 61% 52% 55% 55%
3 https://huggingface.co/transformers/model 𝑜𝑐/𝑑𝑖𝑠𝑡𝑖𝑙𝑏𝑒𝑟𝑡.ℎ𝑡𝑚𝑙
𝑑
3.3.4. Proposed Model Task B
This uses the DistilBERT model trained for Subtask-A. The motive behind using this model is
it’s accuracy on the validation dataset and the model’s reliability. The HOF comments are to be
further classified as PRFN, OFFN or HATE comments.
There are three seperate models combined to form the proposed model. The models are as
follows:
1. The DistilBERT model trained for Subtask-A identifies if a comment is of hate or not.
This is essentially a binary classification (HOF/NOT).
2. Two python libraries ‘profanityfilter’4 and ‘better-profanity’5 are together used to classify
the hate comments from the above model (1) into PRFN or other hate comments.
3. Finally, another DistilBERT model, trained on the OFFN and HATE comments from the
sample dataset, is used to classify the remaining comments respectively.
The flow of the model is as follows: The test data is initially passed to DistilBERT model 1.
above, which is a binary classification model that returns if the given text is hate or not. The
non hate comments predicted by this model are labelled as NONE. And the hate comments are
to be further classified as HATE, PRFN and OFFN.
Now the hate comments in the test data as filtered by the first model as a part of the proposed
model alone is passed to two python libraries ‘profanityfilter’ 4 and ‘better-profanity’ 5 that
classify the hate comments into PRFN hate comments and other hate comments (HATE and
OFFN) to be further classified. Lastly, the other hate comments in the test dataset is passed to a
DistilBERT model trained to classify a text as either HATE or OFFN (trained using the training
dataset provided), which will further classify the text as OFFN and HATE and ultimately, all the
text in the test dataset are now classified as HATE,OFFN,PRFN or NONE.
Fig 9: Proposed Model Block Diagram
Table 6: Validation accuracy results for each model used in our proposed model:
Model Accuracy Macro-F1 Macro-Precision Macro-Recall
1. DistilBERT Subtask-A 78% 73% 75% 72%
2. Profanity filter +
92% 48% 50% 46%
better-profanity
3. DistilBERT for
79% 79% 79% 79%
OFFN/HATE classification
4 https://pypi.org/project/profanity-filter/
Table 7: Overall validation accuracy for our proposed model:
Accuracy Macro-F1 Macro-Precision Macro-Recall
63% 59% 61% 58%
4. Result
The ML model’s accuracy improved using K-Folding Technique with 10 splitting iterations in
the cross-validator. Even though the validation accuracy of the ML models were decently high,
the pre-trained DistilBERT model proved to be most accurate on the test data on submission
with a good 77.67% accuracy score for Sub-Task A and around 65% accuracy for Sub-Task B. Our
proposed model gave the next best prediction for Sub-Task B with approximately 60% accuracy.
The ML models did not perform well when for the Sub-Task B, a possible reason for it could be
that the model could not differentiate between profane, hateful and offensive posts properly.
Data augmentation, exploring other pre-trained models such as XLNet, ERNIE, RoBERTa, and
considering POS tags combined with n-grams to give an extra set of feature space could be the
scope of improvement to solve this problem.
Fig 10: Test Data Accuracies for Sub-Task A Fig 11: Test Data Accuracies for Sub-Task B
Table 8: Results obtained for Sub-Task A
Model Accuracy Macro-F1 Macro-Precision Macro-Recall
DistilBERT 77.67% 74.91% 77.30% 74.91%
LSTM 50.35% 45.08% 45.06% 45.48%
Random Forest 47.77% 47.54% 52.86% 52.61%
SVM 52.06% 49.96% 50.04% 50.05%
Logistic Regression 50.66% 48.13% 48.18% 48.14%
5 https://pypi.org/project/better-profanity/
Table 9: Results obtained for Sub-Task B
Model Accuracy Macro-F1 Macro-Precision Macro-Recall
DistilBERT 65.1% 57.65% 59.90% 57.65%
Proposed Model 59.95% 53.04% 54.98% 53.91%
ML Overall Mode 30.44% 24.25% 25.57% 24.83%
LSTM 35.75% 26.34% 28.66% 27.98%
KNN 32.86% 18.58% 18.58% | 23.19%
Fig 12: Sub-Task A: Confusion Matrix DistilBERT
Fig 13: Sub-Task B: Confusion Matrix DistilBERT
Fig 14: Sub-Task B: Confusion Matrix Proposed Model
5. Conclusion
In this paper, several machine learning and deep learning approaches have been used for
detecting hate speech and offensive language content and the models have been compared.
Several attempts on various techniques to increase the accuracy have been employed. Our
proposed model achieved good results compared to its simplicity, second only to the pretrained
DistilBERT model. We believe with proper feature extraction and data augmentation techniques,
we will be able to improvise our proposed model.
References
[1] Victor Sanh, Lysandre Debut, Julien Chaumond, Thomas Wolf, Distilbert, a distilled version
of bert: smaller, faster, cheaper and lighter, 2020. URL: https://arxiv.org/pdf/1910.01108.pdf.
[2] Apurva Parikh , Harsh Desai , and Abhimanyu Singh Bisht, DA Master at HASOC 2019:
Identification of Hate Speech using Machine Learning and Deep Learning approaches for
social media post, 2019. URL: http://ceur-ws.org/Vol-2517/T3-18.pdf.
[3] Vandan Mujadia, Pruthwik Mishra, Dipti Misra Sharma, IIIT-Hyderabad at HASOC 2019:
Hate Speech Detection, 2019. URL: http://ceur-ws.org/Vol-2517/T3-12.pdf.
[4] MoungHo Yi, Myung, Jin Lim, Hoon Ko, and JuHyun Shin, Method of Profanity Detection
Using Word Embedding and LSTM, 2021. URL: https://doi.org/10.1155/2021/6654029.
[5] H. A. Nayel, S. H. L., DEEP at HASOC2019 : A Machine Learning Framework for Hate
Speech and Offensive Language Detection, 2019. URL: http://ceur-ws.org/Vol-2517/T3-21.
pdf.
[6] Jean-Christophe Mensonides, Pierre-Antoine Jean, Andon Tchechmedjiev, and S´ebastien
Harispe, IMT Mines Ales at HASOC 2019: Automatic Hate Speech Detection, 2019. URL:
http://ceur-ws.org/Vol-2517/T3-13.pdf.
[7] 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.
[8] T. Mandl, S. Modha, G.K. Shahi, H. Madhu, S. Satapara, P. Majumder, J. Schäfer, 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/.
[9] S. Jaki, T. De Smedt, M. Gwóźdź, R. Panchal, A. Rossa, G. De Pauw, Online hatred of
women in the Incels. me forum: Linguistic analysis and automatic detection. Journal of
Language Ag-gression and Conflict, 7(2), 240-268., 2019. URL: http://www.organisms.be/
downloads/incels.pdf.
[10] W. Yin , A. Zubiaga, Towards generalisable hate speech detection: a review on obstacles
and solutions, 2021. URL: https://doi.org/10.7717/peerj-cs.598.
[11] S. Modha, T. Mandl, P. Majumder D. Patel, Tracking Hate in Social Media: Evalua-
tion, Challenges and Approaches, 2020. URL: https://link.springer.com/article/10.1007/
s42979-020-0082-0.