Bengali Hate Speech Detection Using Deep Learning Technique�Team Name: Chandan Senapati�(Paper ID: 75)�HASOC (2023), Task 4�Chandan Senapati, Utpal Roy(Visva-Bharati University, W.B.)

​

​

FIRE 2023, GOA,15-18 December 2023

Outline

1. Hate Speech

2. Bengali Hate Speech in Social Media Platform

3. Importance of Detecting Hate Speech

4. Existing works

5. Our approach

6. Result and Scope of Improvement

7. Conclusion

Hate Speech

• According to Oxford Dictionary abusive or threatening speech or writing that expresses prejudice on the basis of ethnicity, religion, sexual orientation, or similar grounds is called Hate Speech.
• Offensive, discriminatory or harmful words, sentences, emojis, animations signifying hate

​

​

�Bengali Hate Speech in Social Media Platform�

• Bengali hate speech written in Bengali script over social network
• Bengali hate speech written in other script
• Abusive words in Mixed languages

​

Importance of detecting Hate Speech

​

• It poses a significant threat to social cohesion, balance and the mental well-being of internet users
• It undermines social equality as it reaffirms historical marginalization oppression & discrimination.
• It is enacted to cause psychological and physical harm to its victims as it incites violence.
• Indian Information Technology (IT) Act: It regulates online speech, including hate speech. Under the act, intermediaries such as social media platforms are required to remove content that is in violation of the law within 36 hours of being notified.

Challenges to Hate speech:

​

​

​

• The vast majority of hate speech takes place online, and it can be difficult to identify and remove this content. Platforms like Facebook and Twitter have struggled to effectively moderate hate speech, and there is no consensus on how to approach this problem.

​

• India has a diverse population with different languages, religions, and cultures, thus there is a need to curb incidents of hate speech and crimes that can have a detrimental impact on individuals and communities.

​

Existing Works

• Machine Learning and Deep Learning based approaches has been used for different languages.
• Some works have been done in Bengali language using machine learning and deep learning.

Our Approach

​

• We proposed to classify the text using Long Short Term Memory (LSTM) model.
• Deep learning techniques, including LSTM and its variants, have proven to be effective for hate speech detection due to their ability to capture context and sequential patterns in text.

​

�Our Approach�

​

tokenizing, removing stop words, symbols, URLs, stemming, label encoding, etc. were done on the Bengali text data before training the model

​

​

​

Our Approach

• Model Development:
• We harness the capabilities of LSTM networks to build a robust hate speech detection model for the Bengali language. Fine-tuning the LSTM architecture, optimizing hyperparameters,and experimenting with word embeddings and padding we try to maximize the accuracy and efficiency of our system. For that, We divided the data set into training and testing keeping a ratio of 80:20

Result

• The HASOC (2023) team assessed the results using the Macro F1 score as the performance metric. The score was 0.50625.

​

• Converting emojis and emoticons to text helps to increase performance. More experiments on preprocessing of Bengali text are needed to increase the model’s performance.

​

• Experimenting on different NLP models is necessary to compare performance

​

​

Conclusion

• This research tries to contribute to mitigating the harmful impact of Bengali hate speech used in social networks and to provide a safer and more inclusive online environment for Bengali-speaking communities, by developing a reliable and efficient deep learning LSTM network.
• We also explore potential applications of our model, including content moderation on social media platforms, early detection of hate speech trends, and support for online safety initiatives.

​

​

Thank You