By NUSH Team 0 (Khang, Yuan Xi, Edden, Prannaya, Yu Pin and Yaw Tia)
Made for NTU Deep Learning Week Hackathon 2022
30th September 2022 to 3rd October 2022
With the impending rise of the metaverse, and with social media and more permeating modern culture in a way never before seen, children have become more likely to interact with others on the internet. Due to hyperstimulation, kids are more likely to venture into using technologies like Roblox and Youtube, and that leads to rampant problems in their health and their ability to focus.
Although Roblox appears to be a harmless game where kids can use their creativity, there is also a lot of adult content inside the game that is inappropriate for children. In addition, such platforms are unable to cater to the necessary content moderation and filtration needed to protect our children, as we enter a smart and digital nation. Scamming is also a common phenomenon on the internet, according to this Straits Times article from March 2022, which makes it even more dangerous since children may be urged to click on such links and could experience guilt for having done that soon after. According to this post from Today.com, an individual was essentially assaulted in this platform, and the only reason what that there was no proper content moderation. Hence, it is incredibly crucial that we provide a utility to parents and children for an additional layer of well-needed safety before such content reaches them.
Similar to the problem once identified for spam classification problems, we must be able to account for the slight variations in wording, wherein often pages can be clouded with numerous incorrect spellings of the profanities intended, which don't get removed from the pages. f0ck instead of f*** is a common example. While most utilities simply permute these combinations, we have decided to use a novel AI framework to do this. Via Python's pillow library, we convert each word into a separate image, which we pass through a very tiny CNN model in order to determine it's closeness to any one of the closed set of profanities we have compiled as a list. This AI model is able to perform simple similarity detection algorithms, although it is still susceptible to specific spellings, like fu0k, failing to register as similar. In addition, it can sometimes detect texts which are not meant as profanities, such as singapore and pofma, and simply classify it as a vulgarity. In such cases, an automatic bypass database has been created which looks up specific texts and whitelists them.
Last but certainly not least, we have developed an NLP model to identify such problematic texts and which uses sequential sentence classification to remove such texts. Due to the smaller amount of data, we have instead used two datasets:
OpenAI's Moderation API Dataset (1690 Records)
We have used OpenAI's Moderation API as a basis for labelling, as we feel that this task requires a larger range of potential classifications. The Moderation API effectively provides the following scores, as computed by their own AI:Category Label Definition sexual SContent meant to arouse sexual excitement, such as the description of sexual activity, or that promotes sexual services (excluding sex education and wellness). hate HContent that expresses, incites, or promotes hate based on race, gender, ethnicity, religion, nationality, sexual orientation, disability status, or caste. violence VContent that promotes or glorifies violence or celebrates the suffering or humiliation of others. harassment HRContent that may be used to torment or annoy individuals in real life, or make harassment more likely to occur. self-harm SHContent that promotes, encourages, or depicts acts of self-harm, such as suicide, cutting, and eating disorders. sexual/minors S3Sexual content that includes an individual who is under 18 years old. hate/threatening H2Hateful content that also includes violence or serious harm towards the targeted group. violence/graphic V2Violent content that depicts death, violence, or serious physical injury in extreme graphic detail. Profanity Check's Combined Binary Classification Dataset (180k records)
This Dataset only contains binary labels of Profanity detection, but we utilise it as a larger unlabelled data source with some confirmation for pseudolabels.
We then create a simplistic "Teacher" model, a small Bidirectional LSTM as trained via TensorFlow on Kaggle's Notebook feature. This is a small model which uses the advantage of using the smaller first data sample to train over a small period of time. After this, we pseudolabel the second dataset and verify the results. Following this, with
For the AI to be incorporated into our application, we use a Flask server which has been deployed on Anaconda's PythonAnywhere service. We are then able to run API request for the batch of text in both of the following utilties.
The PythonAnywhere utility, allowing us to openly interface with the AI models trained.
The surge.sh functionality, as per the domain.
While we originally planned on a specific roster, this kept shifting and we ended up with more on our plate than we could handle. Some of us were also involved in other events, such as the Tech4Good Festival and generally other events. However, we still pulled through.
We have not just developed, but deployed our servers and clients online for people to be able to use, which we are incredibly proud of. We are also proud of using several innovative solutions with little to no prior literature, which sets up apart from other projects.
Always start with a solid plan.
Despite this obvious fact, it truly came home with this project, where we went into the coding process with well-established expectations and general expectations, which allowed us to develop more and more technologies and change a small amount of the original app, without it failing much of the rest.
"Develop AI models or IoT solutions that solve industrial or social problems in the new stage of society development. The Smart Nation is an initiative by the Government of Singapore to harness from info-comm tech, networks and big data to create tech-enabled solutions."
| Round | Time | Deliverables |
|---|---|---|
| 0 (Submission) | Day 4, 6am | - 8 Slides Deck - 30s Demo Video - Source Code on GitHub |
| Round 1 | Day 4, 9am to 1.30pm | 30 Teams selected for 5 minute demo |
| Round 2 | Day 4, 2pm to 5.30pm | 4 min Presentation, followed by 6 min Q&A |
| Award Ceremony | Day 4, 6pm |
| Task | Subtask | Percentage |
|---|---|---|
| Technical Difficulty | Execution, Task Difficulty | 40% |
| Industrial Value | Feasibility, Budget Cost, Community Impact, Business Value | 30% |
| Innovation (Idea) | Breakthrough, Idealization | 20% |
| Project Delivery | Presentation, Readability, UI/UX | 10% |