Hate speech and online violence are rampant nowadays, SafeMedia aims to make social media a safe place for everyone!
While the 2019 General election is approaching in our country Nigeria, one of the most mitigating factors against the peaceful co-existence of the nation of different tribes and multiple religions and aggravate our democratic governance is hate speech and online violence.
Hate speech and online violence result in many numbers of crises in Nigeria that claims thousands of lives and properties worth billions of dollars across its regions. Different attempts have been made by Government and nongovernmental organizations to mitigate if not eliminate the act among its citizens hence the Nigerian Government passed a bill that makes hate speech an act of terrorism.
Fortunately, Nigeria is a diverse country that comprises hundreds of tribes and numerous religious thoughts and beliefs these factors results in natural disharmony among the citizens, when someone said something according to his belief and thoughts others might feel offended.
Our research aims to bring communities together promoting nonuse of any form of hate speech in social media and to assist users in analyzing what to write and what to read that will not offend others. Also to provide a public API for developers, researchers and security concerns on hate speech and online violence.
What it does
SafeMedia request user's Facebook permission to post and read data on his behalf. Safemedia helps a user to write a non-hate speech on his Facebook wall and also read his group post safely and hide any hate speech post. If a user feels any message or post contain hate speech and the system didn't recognize it, a user can instantly report that message or post as a hate speech.
How We built it
We collected and gathered more than 10,000+ social media post from various social media users of the two most spoken languages in Nigeria (Hausa and English) that include both hate and non-hate speech. Morealso, we build a sentiment analysis algorithm using ML.NET machine learning framework with 83% accuracy to predict hate speech in the two languages.
Challenges We ran into
Due to recent data privacy laws that enforce social media giants change their data sharing policy, we find it difficult to find a tool that will generate the data to train our model. Most of the public data we got is not the target language (Hausa) hence we collect more than 5,000+ Hausa post manually from social media platforms. Also, the initial schema we design to work with has so many data inputs to the extent that to fill a single row it will take you more than 5 minutes, after all, we design a simple schema that we use to collect more than 6,000+ data manually from social media.
Accomplishments that We are proud of
- We built an AI model that has 99% language prediction
- Also, our Hate Speech analysis model has 83.2% accuracy and supports both 2 most spoken language in our community (Nigeria); Hausa and English
- The first Hausa sentiment analysis AI model
- The first time for us to build an AI model serving on production
- Potentially bringing a community of peace lovers together by promoting non-use of hate speech on social media
What we learned
Being our first time to build AI model serving on production, we have learned a lot on model training and serving on a production scale.
What's next for SafeMedia
Our next iteration in the roadmap is to support other Nigerian languages such as
- Kanuri Also to support other social media platforms like
- Google Plus Moreso, to implement the whole features as a browser extension
- ML.NET (Machine Learning framework)
- Facebook Login SDK
- Facebook Graph API