HateCheck

An application to detect Hate speeches in social media

"Darkness cannot drive out darkness; only light can do that. Hate 

      cannot drive out hate; only love can do that"- Martin Luther King Jr.

These words of Martin Luther King still echo in our minds when we hear of incidents such as Riots in Bangalore incited due to Facebook post, US Capitol Hill Siege, or more recently, the Shaheen Bagh Protest where peaceful protestors were labeled 'anti-nationals' and 'jihadists.' These incidents have one thing in common. They were all incited by or at least propagated by organized hate speech campaigns especially through social media as a tool. This brings to fore the immense potentialities of social media in propagating Hate Speeches. Given its wide relevance in today's world, it becomes crucial to understand what ‘Hate Speech’ is and how it can be prevented?

United Nations (UN) defines hate speech as "any kind of communication in speech, writing or behavior, that attacks or uses pejorative or discriminatory language with reference to the person or a group on the basis of who they are, in other words, based on their religion, ethnicity, nationality, race, color, descent, gender, or other identity factors". In India, it is manifested mainly on religious and caste lines. The consequences of hate speech can be devastating. It has the potential to breed intolerance and extremism in society. This was well understood by dictators of the likes of Hitler who used intense and consistent propaganda against Jews to create an entrenched opinion amongst the masses which ultimately culminated into widespread persecution of Jews, remembered in history as the Holocaust.

Now that we know the potential of Hate Speech, it becomes imperative to prevent it. How can we prevent hate speech?

I, along with my friends, have come up with a solution as a part of our project 'Hate-Check' for a course 'Privacy and Security in Online Social Media (PSOSM)' we took this semester in IIIT Delhi. We have tried to address the problem by creating a website that takes the URL of a tweet from the users and identifies whether it contains hate speech or not. If the tweet contains traces of hate speech in it, the website identifies the section that it has flagged as containing hate speech. It also classifies hate speech into 6 different categories viz., Not Hate, Racist, Sexist, Homophobic, Religious, Other. It will help users become aware of such tweets online and make informed choices while consuming the content.

 There can be a multitude of uses of ‘HateCheck.’ Firstly, through this application, a user  can save himself from the toxic effect of consuming such information at an individual level. At a community level, it can help by reducing hate comments online by alerting users before posting anything that can intentionally or unintentionally cause hate. It can fast-track judicial processes by acting as a litmus test to objectively determine whether a tweet can be considered as hate speech or not.

How does it work?

The Machine Learning model has been used to implement the solution. A training dataset consisting of labeled texts and reviews annotated on the labels by content reviewers are used to building a Machine Learning classifier that can subsequently, be used to detect hate speeches in a text. In Layman's terms, our classifier would learn from the training dataset fed to it and would use that algorithm to classify hate speeches.

Methodology

Our methodology is based on four key steps, namely Data Collection, Data. Preprocessing, Constructing Classifier Model, and Classification Model Evaluation. 

Let us understand each step in a little more detail

Data Collection

The machine-learning classifier has been trained with the help of a dataset consisting of more than 1.46 lakh tweets. Each tweet has been labeled by a content reviewer as one of the following labels:

1. Not Hate
2. Racist
3. Sexist
4. Homophobic
5. Religious
6. Other 

Data Pre-processing

In this stage, the data collected has to be organized in a manner such that data-processing steps can be applied to it. For this, we remove the following features if they are present in the tweet:

1. Images: Images cannot be processed through our algorithm.
2. Special Characters: Special characters like !,#, @, (, ), etc. if present are removed as they cannot make a tweet more or less hateful
3.  Stop words: The words like I, We, They, are, our, All, At, are, etc. are removed

The remaining words are then converted into shingles. This is to ensure that each word of the text is processed in the later steps

Constructing Classifier Model

At this step, a machine learning classifier is trained using the pre-processed data. To train the model, we need to associate an integer value as an identifier for each of the data labels. Following identifiers have been used:

0- Not Hate

1- Racist

2-Sexist

3-Homophobic

4-Religious

5-Other

We trained our model using the following machine learning models:

1. Gaussian Naïve Bayes (GNB)
2. Support Vector Machines (SVM)
3. Decision Tree (DT)
4. K-Nearest Neighbours (KNN)
5. Random Forest

Classification Model Evaluation

Each of the trained models are evaluated objectively by considering their precision and recall values for a few tweets. It has been observed that the Random Forest The classifier model had the highest precision and a very high value of recall. Hence it can be inferred that ‘Random Forest Classifier’ is the best-trained model for our dataset among all models.

HOW TO USE OUR WEBSITE

Step 1: Go to our website http://127.0.0.1:8000/

Step 2: You will see the description of our project on the website. Scroll down to enter the URL of the tweet which you want to check

Step 3: You will get the result that our model predicted, i.e., whether the tweet is hateful or not.

Conclusion

Hate Speech is an emerging problem of the 21st century whose effects are only going to increase. Hence there is an urgent need to monitor and regulate hate speeches as they breed extremist views and intolerance. HateCheck is a step in the direction that can reduce the hateful comments in social media by alerting users before posting if their content is hateful and by highlighting hateful content in the text. In the future, we would like to explore the idea of identifying hate speeches in different media like audio and video. 

Team Members

Vinay Dayal (2018117)    Shivam (2018309)     Kaiman Yadav(2018288)

 Yash Kanojia (2018372)      Arbaaaz Khan(2018023)    Sandeep Sharma(2018093)

Ankit Kumar(2018018)     Shubham Manjhi(2018194)    Avinash(2018227)

Bhaskar(2018332)        Nikhil Yadav(2018298)     Gauri Shankar(2018035)