DEEP at HASOC2019 : A Machine Learning Framework for Hate Speech and O ensive Language Detection

Abstract

In this paper, we describe the system submitted by our team for Hate Speech and O ensive Content Identi cation in Indo-European Languages (HASOC) shared task held at FIRE 2019. Hate speech and o ensive language detection have become an important task due to the overwhelming usage of social media platforms in our daily life. This task has been applied for three languages namely, English, Germany and Hindi. The proposed model uses classical machine learning approaches to create classi ers that are used to classify the given post according to di erent subtasks.

Keywords

• Multi-task Classi cation
• Multi-lingual Text Analysis
• Hate
• Speech and O ensive Detection 3

-

Wide spread use of internet is giving more freedom to people to express their thoughts freely and anonymously in di erent forms such as blogs, business networks, forums and social media such as Twitter and Facebook. Internet has enabled the interaction among people from di erent culture, race, religion, origin, gender and nationality. It has also paved the way for the online information to reach the mass within a matter of seconds. Social media is generating lot of content as it has opened up a new galaxy of opportunities for people to express their opinions online resulting in the exchange of ideas in a positive way as well as contributing to the propagation of hate speech in a negative way. The anonymity of users provided by the internet is also impacting the society in generating the web contents in a negative way by the usage of profane words, hate speech, 3 Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). FIRE 2019, 12-15 December 2019, Kolkata, India. derogatory terms, racists slurs, toxic, abusive and o ensive language. This negative content may be aggressive and potentially harmful lowering the self-esteem of people leading to mental illness and suicidal attempts and has also forced people to deactivate their social media accounts. The increase of cyberbullying and cyberterrorism, and the use of hate speech content on the Internet, make the identi cation of hate speech an essential ingredient for anti-bullying policies of social media [ 5 ]. Many terrorist activities, which are related to hate speech, are gaining huge attention in social media in terms of posts and suggestions [ 1 ].

Hate speech and other o ensive and objectionable content online are posing threats and challenges to the society. A lot of countries prohibit hate speech in social media subject to the condition that it should not target any group or trigger any crime. As hate speech acts as a opinion builder, many online forums such as YouTube, Face book and Twitter, have their own policies to remove hate speech content or anything which is impacting the society in a negative way.

It is therefore important to take preventive measures to cope up with hate speech and objectionable content or remove such content. Detecting hate speech is a di cult task as there is no clear de nition for hate speech and can vary according to the context. For example, speech which contain subtle and nuance sentences are di cult to detect as hate speech. Further, the usage of slang terms may be found neutral, but may trigger hate crime or o ense in some context. Although there is no universal de nition for hate speech, the most accepted de nition is provided by Nockleby [ 12 ]: \any communication that disparages a target group of people based on some characteristic such as race, colour, ethnicity, gender, sexual orientation, nationality, religion, or other characteristic" [ 5 ].

Detection and removal of hate speech and objectionable content manually is a tedious task due to the massiveness of the web and the increasing number of online users. Further, the task becomes di cult due to the anonymity of users on the internet. Hence, there is a great demand for tools and techniques that automatically detect the hate speech and objectionable content quickly on the web to reduce the spread of hate speech.

In this paper, we describe the system submitted by our team for the shared task of Hate Speech and O ensive Content Identi cation in Indo-European Languages (HASOC) track at FIRE 2019 [ 7 ]. Rest of the paper is organized as follows: In Section 2 we review the related work on hate speech and o ensive language detection. Description of the task and a summary of dataset is given in Section 3. Our proposed model is discussed in Section 4 followed by the experiments and results in Section 5. Finally, we conclude the paper in Section 6. 2

Related work

Though, de ning and understanding hate speech is di cult, several online forums, IT industries and researchers have explored many algorithms to detect hate speech, o ensive and abusive content on the web [ 5, 2, 6, 3, 15, 4 ]. Thomas et.al. [ 2 ] trained a multi-class classi er to classify tweets into one of three categories, namely, hate speech, o ensive but not hate speech, neither o ensive and nor hate speech. They performed experiments using logistic regression, Nave Bayes, decision trees, random forests, and linear SVM classi ers with 5-fold cross validation and obtained an overall precision 0.91, recall of 0.90, and F1 score of 0.90 for the best performing model. Gibert et.al. [ 5 ], constructed a manually labeled hate speech dataset composed of thousands of sentences obtained from Stormfront, a white supremacist online forum which contains hateful and no hateful sentences. Their model was built using Support Vector Machines (SVM), Convolutional Neural Networks (CNN), Recurrent Neural Networks with Long Short term Memories to annotate the test data.

Sean MacAvaney et.al. [ 6 ], examines the challenges faced by online automatic approaches for hate speech detection in text in addition to the existing approaches and various datasets for hate speech detection. They also have proposed a multi-view SVM approach that achieves near state-of-the-art performance, while being simpler and producing more easily interpretable decisions than neural methods. The experiments were conducted on HatebaseTwitter, Stanfront and TRAC datasets. Avishek Garain and Arpan Basu [ 3 ], presents a description of their system to detect o ensive language in Twitter submitted to \SemEval-2019 Task 6 [ 14 ]" which uses bidirectional LSTM, a neural network based model to capture information from both the past and future context. Ziqi and Lei [ 15 ], discusses about the typical features responsible for classi cation between hate speech and no hate content. In their work they proposed Deep Neural Network structures serving as feature extractors that are particularly e ective for capturing the semantics of hate speech. Their methods were evaluated on the largest collection of hate speech datasets based on Twitter, and are shown to be able to outperform the best performing method by up to 5 percentage points in macro-average F1, or 8 percentage points in the more challenging case of identifying hateful content. Aditya et.al., [ 4 ] proposes an approach to automatically classify tweets on Twitter into three classes: hateful, o ensive and clean by considering features like n Grams and TF/IDF values. They performed a comparative analysis of the models considering several values of n in n-grams and TF/IDF normalization methods on Twitter dataset and achieved 95.6% accuracy. Areej Al-Hassan and Hmood Al-Dossari [ 1 ] in their survey paper, gives detailed information about hate speech in terms of insu cient dataset, complexities and challenges and the complete knowledge about text mining and Natural Language Processing techniques to nd and classify hateful and no-hateful contents. 3

Task Description and Corpus

Hate Speech and O ensive Content Identi cation in Indo-European Languages (HASOC) shared task aims at identifying the hate speech and o ensive content in three Indo-European languages namely English, Germany and code-mixed Hindi. The shared task is divided into three sub-tasks to identify the hate speech and o ensive content posts and classify them into various prede ned categories. The rst sub-task, Sub-task A is a coarse-grained binary classi cation that focuses on the identi cation of hate speech and o ensive language in the given posts and is o ered for English, German and Hindi language posts. Sub-task B is a ne-grained multi-class classi cation that classi es the hate speech and o ensive language posts into one of the three prede ned categories, namely hate speech, o ensive and profane. This subtask is o ered for English, German and Hindi language posts. The third sub-task, Sub-task C aims at checking the type of o ense as targeted insult or un-targeted insult and is o ered for English and Hindi. The corpus consists of posts collected from Twitter tweets and Facebook comments in English (EN), German (GR) and Hindi (HI) languages. Each post contains three labels for Sub-task A, Sub-task B and Sub-task C respectively. Table 1 gives a glimpse of the shared task and the associated categories in each sub-task. A detailed description of our model and the classi ers used are given in this section. Given a set of posts P = fp1; p2; :::; png, where each post is composed of a set of words pi = fw1; w2; :::; wkg, in each language L = fEN; GR; HIg and the classes A = fN OT; HOF g, B = fHAT E; OF F N; P RF N g and C = fT IN; U N T g for the sub-tasks A, B and C respectively, the shared task is to formulate a multilabel classi cation problem, where an unlabeled post/instance is assigned with multiple class labels one from each class A, B and C. ie., given a unlabeled post pk, multi-label classi cation problem will assign the triple < a; b; c > such that, a 2 A, b 2 B and c 2 C. 4.2

Model The general structure of the proposed model is given in Fig. 1. This model creates sub-models for each subtask and then assigns multiple class labels, one for each subtask, by combining the output of each sub-model. The model consists of the following phases:

I. Preprocessing Preprocessing also known as corpus cleaning is a crucial and rst step in building any classi ers. Each post pk has been tokenized into a set of words to get n-gram (n = 2) bag of words. In this step, all un-informative tokens such as urls, digits and special characters have been removed from all the posts.

II. Feature Extraction In this phase, a TF/IDF vector has been computed for all the posts in the training set. This vector will be used as an input for training the classi er. TF/IDF has been calculated as described in [ 10 ]. III. Training the Classi er In this phase, the features that have been extracted in previous phase are used as input for training the classi ers. Three classi cation algorithms namely, Linear classi er, SVM and Multilayer Perceptron (MLP) have been used separately to train the proposed model. Linear classi er is a simple classi er that uses a set of linear discriminant functions to distinguish between di erent classes [ 13 ]. SVM is a kind of linear classi er which uses the samples close to the classes's boundaries for training and these samples are known as support vectors [ 11 ]. SVM has been used e ectively in di erent NLP tasks [ 9, 8 ]. MLP is a deep learning approach that uses back propagation for training the neural network [ 10 ]. It is characterized by several layers of input nodes connected as a directed graph between the input and output layers.

Three di erent classi ers namely, Linear classi er, SVM and MLP are created for each subtask. Classi ers for Sub-task A will classify the unlabeled instance into one of the two prede ned categories as mentioned in Table-1. While the entire data is used to construct classi ers for Sub-task A, a sub-corpus (only posts labelled as HOF in Sub-task A) that contain only the hate speech and o ensive language posts is used as training corpus to create classi ers for Sub-tasks B and C. Classi ers for Sub-task B will classify the unlabeled instance into one of the three prede ned categories and classi er for Sub-task C will classify the same instance into one of the two prede ned categories as mentioned in Table 1. 5

Experiments and Results

In our proposed model, the Stochastic Gradient Descent (SGD) optimization algorithm has been used for optimizing the parameters of linear classi er, while SVM uses Lagrange multipliers to solve the optimization problem [ 13 ]. The loss function used in linear classi er was "Hinge" loss function. Linear kernel has been used for SVM classi er. In MLP classi er the logistic function has been used as activation function using 20 neurons in the hidden layer.

We have used cross-validation approach with ve folds to train di erent submodels for each subtask and the outputs of each sub-task have been combined to generate the nal output. The organizers of the shared task used Macro-averaged F1-score (M-f1) and weighted F1-score (W-f1) [ 10 ] as performance evaluation metrics for all sub-tasks. The dataset comprises of a set of tweets and Facebook comments in the three languages labelled with di erent categories. Statistics of the dataset is shown in Table 2.

Table 3 shows the M-f1 and W-f1 of our submissions for all sub-tasks over English, Germany and Hindi language posts. From the table, it is clear that SVM outperforms other classi ers for all subtasks for English. For German language posts, MLP outperforms other classi ers for Sub-tasks A and B. It is worth to note that, Sub-task C is not applied for Germany. For Hindi language posts, MLP outperforms other classi ers for Subtask A, while SVM reported better results than MLP and linear classi er for subtask B.

There is a gab between M-f1 and W-f1 for all subtasks except for the subtask A for Hindi. This is due to the fact that the performance of the subtasks B and C depends on the performance of the subtask A. If the prediction of subtask A is wrong then by default subtasks B and C will also have wrong prediction. Sub-task Class Labels English German Hindi In this paper, a machine learning approaches have been used for creating a model for detecting hate speech and o ensive language content. Proposed model achieved good results compared to its simplicity. Extension of our work includes using deep learning approach to build the classi er and test it on much bigger dataset.

References

1. Al-Hassan , A. , Al-Dossari , H.: Detection of Hate Speech in Social Networks: A Survey on Multilingual Corpus . In: Proceedings of 6th International Conference on Computer Science and Information Technology (CoSIT 2019 ). pp. 83 { 100 . Dubai , UAE (February 23 -24 2019 )

2. Davidson , T. , Warmsley , D. , Macy , M.W. , Weber , I. : Automated Hate Speech Detection and the Problem of O ensive Language . In: Proceedings of the Eleventh International Conference on Web and Social Media , ICWSM 2017 , Montreal, Quebec, Canada, May 15 -18, 2017 . pp. 512 { 515 . AAAI Press ( 2017 )

3. Garain , A. , Basu , A. : The Titans at SemEval-2019 Task 6: O ensive Language Identi cation, Categorization and Target Identi cation . In: May, J. , Shutova , E. , Herbelot , A. , Zhu , X. , Apidianaki , M. , Mohammad , S.M. (eds.) Proceedings of the 13th International Workshop on Semantic Evaluation, SemEval@NAACL-HLT 2019 , Minneapolis , MN, USA, June 6-7, 2019 . pp. 759 { 762 . Association for Computational Linguistics ( 2019 ), https://www.aclweb.org/anthology/S19-2133/

4. Gaydhani , A. , Doma , V. , Kendre , S. , Bhagwat , L. : Detecting Hate Speech and O ensive Language on Twitter using Machine Learning: An N-gram and TFIDF based Approach . CoRR ( 2018 ), http://arxiv.org/abs/ 1809 .08651

5. de Gibert , O. , Perez , N. , Garc a-Pablos, A. , Cuadros , M. : Hate Speech Dataset from a White Supremacy Forum . In: Proceedings of the 2nd Workshop on Abusive Language Online (ALW2) . pp. 11 { 20 . Association for Computational Linguistics, Brussels, Belgium (Oct 2018 ). https://doi.org/10.18653/v1/ W18 -5102

6. MacAvaney , S. , Yao , H.R. , Yang , E. , Russell , K. , Goharian , N. , Frieder , O. : Hate speech detection: Challenges and solutions . PLoS ONE ( 2019 )

7. Modha , S. , Mandl , T. , Majumder , P. , Patel , D. : Overview of the HASOC track at FIRE 2019: Hate Speech and O ensive Content Identi cation in Indo-European Languages . In: Proceedings of the 11th annual meeting of the Forum for Information Retrieval Evaluation (December 2019 )

8. Nayel , H. , Shashirekha , H.L. : Improving NER for Clinical Texts by Ensemble Approach using Segment Representations . In: Proceedings of the 14th International Conference on Natural Language Processing (ICON-2017) . pp. 197 { 204 . NLP Association of India, Kolkata, India (December 2017 ), http://www.aclweb.org/anthology/W/W17/W17-7525

9. Nayel , H.A. : NAYEL@APDA: Machine Learning Approach for Author Pro ling and Deception Detection in Arabic Texts . In: Proceedings of the 11th annual meeting of the Forum for Information Retrieval Evaluation (December 2019 )

10. Nayel , H.A. , Shashirekha , H.L. : Mangalore University INLI@FIRE2018: Arti cial Neural Network and Ensemble based Models for INLI . In: Mehta, P. , Rosso , P. , Majumder , P. , Mitra , M. (eds.) Working Notes of FIRE 2018 - Forum for Information Retrieval Evaluation, Gandhinagar , India, December 6- 9 , 2018 . CEUR Workshop Proceedings , vol. 2266 , pp. 110 { 118 . CEUR-WS.org ( 2018 ), http://ceurws.org/Vol- 2266 / T2 -10.pdf

11. Nayel , H.A. , Shashirekha , H.L. , Shindo , H. , Matsumoto , Y. : Improving MultiWord Entity Recognition for Biomedical Texts . CoRR abs/ 1908 .05691 ( 2019 ), http://arxiv.org/abs/ 1908 .05691

12. Nockleby , J.T.: Hate Speech . In: Encyclopedia of the American Constitution . pp. 1277 { 1279 ( 2000 )

13. Theodoridis , S. , Koutroumbas , K. : Chapter 3 - Linear Classi ers . In: Pattern Recognition (Fourth Edition) , pp. 91 { 150 . Academic Press, Boston, fourth edition edn. ( 2009 )

14. Zampieri , M. , Malmasi , S. , Nakov , P. , Rosenthal , S. , Farra , N. , Kumar , R.: SemEval -2019 Task 6: Identifying and Categorizing O ensive Language in Social Media (O ensEval) . In: Proceedings of the 13th International Workshop on Semantic Evaluation . pp. 75 { 86 ( 2019 )

15. Zhang , Z. , Luo , L. : Hate Speech Detection: A Solved Problem? The Challenging Case of Long Tail on Twitter . CoRR abs/ 1803 .03662 ( 2018 ), http://arxiv.org/abs/ 1803 .03662