Cyberbullying.AI

Cyberbullying is becoming increasingly prevalent on social media platforms like Facebook and Instagram, text messaging apps, and online forums. In fact, in 2018, the Center for Disease Control and Prevention estimated that 15.7% of students in grades 9-12 were bullied online. A year later, in 2019, the National Center for Education Statistics and Bureau of Justice stated that 16% of high schoolers had experienced cyberbullying. On Facebook alone, in 2016, 85% of 19-year-old men claimed to have experienced online threats and 75% of 18-year-old women claimed to have been victims of the same. This is not just a problem regarding teenagers, in 2016, 7% of working adults in the US reported to have been cyberbullied at least once at their workplace. This amounts to 9.8 million people. However, according to the Cyberbullying Research Center, 50.9% of adolescent girls have experienced cyberbullying compared to 37.8% of boys. Additionally, the non-profit, Ruling Our Experiences stated that girls are five times more likely to be depressed using technology. Clearly, girls experience cyberbullying at a greater rate and are victims of a greater impact, which can develop into mental health problems and learning difficulties because bullying can have lifelong effects on the victim’s mental health and confidence. Often, victims demonstrate signs of lacking the motivation to participate socially, having trouble focusing, and losing interest in hobbies. Over time, they develop more severe disorders including depression as it has been recorded that approximately 30% of children who have been cyberbullied have suicidal thoughts. However, it is becoming increasingly common for someone to take advantage of the virtual barrier between themselves and those online to anonymously embarrass, humiliate, and demean others. 

Cyberbullying.AI is an app and virtual keyboard/extension that implements a novel natural language processing and machine learning pipeline to detect and correct several forms of cyberbullying to mitigate the harmful effects of cyberbullying in online forums, social media apps and websites. This app has two major components to catch cyberbullying before harmful content is posted and provide suggestions on how the user can write a more appropriate and mindful message.

1) A cyberbullying detection AI model to identify comments and texts that embarrass, humiliate, and demean others in the forms of racism, sexism, exclusion, harassment, body-shaming, etc with 92% accuracy, which is around most top-performing models. 

2) A cyberbullying correction AI model to provide suggestions to the bully on how they can better phrase their harmful words. 

Cyberbullying correction has not been implemented on online platforms. However, this is a promising solution that can reduce the rate of cyberbullying considering that just through the implementation of cyberbullying detection, the app "ReThink" reported that offensive messages dropped from 71% to 4%. This app/virtual keyboard will not coerce bullies or users to correct/rephrase their words; however, as illustrated by the "Rethink" app, this method proves effective in education and the workplace and can be even more so with the addition of suggestions from the correction AI model, creating a more welcoming and productive online environment. 

Cyberbullying.AI's primary target customers are social media platforms, online forums, students, educators, and those in the workforce, essentially anyone who has experienced cyberbullying, would like to be more self-conscious when texting, and spends a large portion of their day using digital media.

1) Geographic Segmentation: As stated above, according to  the Center for Disease Control and Prevention estimated that 15.7% of students in grades 9-12 were bullied online, and 9.8 million American adults currently struggle with cyberbullying in the workplace. Both statistics imply a large global pool of customers from children to adults including students, educators, celebrities, and those in the workforce. The novel software primarily targets those who are highly active on social media or online forums including high school and college students, gamers, celebrities, and programmers. The software’s secondary geographic target market will be educators and others in the workforce. 

2) Demographic Segmentation: Demographically, the software’s target audience is not dependent on socioeconomic status as the goal is to reduce cyberbullying and demeaning comments within the digital community. Marketing interest will be shown in all ages from young students to middle-aged workers to the retired elderly. Gender, ethnicity, and religion do not affect the target market as well.

3) Behavioral Segmentation: The software’s primary target customers are likely to be large corporations that would like to promote a healthy and growth-inducing work environment or are individuals who have experienced cyberbullying, would like to be more self-conscious when texting, and spend a large portion of their day engaged in digital media. Additionally, educators and those in the workforce would benefit from having the software to effectively and appropriately communicate with others.

I have noticed the prevalence of cyberbullying online, especially out of school, and am actively receiving feedback from experts in the fields of NLP and ML and those in the workforce through meetings and research conferences. 

As a female adolescent, I can effectively gauge the use of social media and other online platforms in cyberbullying in and out of school and am directly able to communicate with and understand the needs of the target market. Using feedback from computer science and artificial intelligence professors at the University of North Texas, I was able to significantly improve the accuracy of my AI models. I also presented my project at an International Youth Research Conference and was approached by several experts in the field of natural language processing. Many were excited by the idea that we can use AI and natural language processing to prevent cyberbullying and saw it as pragmatic in their own workplace. However, they also suggested that I train on a larger dataset to ensure that cyberbullying correction mechanism can reduce the cyberbullying rate to an even lower percentage than what current detection apps have been able to accomplish. In the future, my goal is to connect with a greater number of experts in the field to improve the model architecture and collect data through crowdsourcing.

Currently, the technology can be visualized using an AI chatbot I trained. The chatbot is a prototype of how the NLP pipeline will be implemented as an extension of current social media and online platforms. Through the chatbot, one can see that when the user is about to send a harmful message, the detection model provides the bully with a popup that lists suggestions from the correction model to help rephrase their words. I have also published and presented my chatbot prototype, model prototypes, and accuracy results in a paper and to professors at the University of North Texas.

I applying to this challenge to gain access to a network of resource partners and experts in the field of natural language processing and machine learning, to learn more about business, and find more resources to improve my technology. Currently, my AI models are overfitting on the small amount of data available and are not generalized enough to go past the prototyping stage. Through the resources offered by this challenge, I wanted to further improve my models by implementing a larger dataset through data augmentation or crowdsourcing. 

As stated before, as a female adolescent, I can effectively gauge the use of social media and other online platforms in cyberbullying in and out of school and am directly able to communicate with and understand the needs of the target market. I have access to and have effectively used the feedback of computer science and artificial intelligence professors at the University of North Texas, those in the workforce, and experts at research conferences over my chatbot and model prototypes.