HSG is an Crome/Edge extension that detects hate speech in real time on web pages.
We hope HSG will have a good influence on children, parents and global.
Hate Speech refers to statements that contain prejudices against certain groups, deliberate slander, threats, and incitement aimed at inciting violence.
Exposure to hate speech is an important factor that can affect the emotions of growing children. Children live in numerous hateful expressions even though they know themselves or without their knowledge. This exposure to hate speech has become even more serious as online activities have increased in the current COVID-19 era.
Exposure to hate speech adversely affects children's emotional development and value formation.
HSG seeks to present with the functions of the next chapter a new method of education that can recognize, prevent, and cope with discrimination in freedom of expression and opinion. When children are defenselessly exposed to hate speech, they recognize it and help them cope with it. HSG filters and removes web text and video and provides them to users through an extension program. By using HSG, it will be possible to minimize the exposure of children's hate speech and help children to form correct values and emotions.
- Edge/Crome Extenstions
- Azure
- TensorFlow
HSG is the Web Extension Program. When you download the program to the web browser you use, it will run automatically detect hate speech.
HSG inspects the text in the web site in real time. If it is determined that some text contains hate speech, HSG closely analyzes the text. The analysis results show the words that contributed greatly to the judgment of hate speech. Using this result, text masking is performed win the web site.
HSG checks the Hate Speech of the video in real time. First, HSG finds the text in the video. If the text is determined to be hate speech, store the found time and duration. Then, video masking conversion within the web site is done using the stored time and duration.
HSG considers the degree of user freedom. The option page helps parents to discuss with their child how to recognize and deal with the hate speech through the selection of users (mute - not mute / display - no notification). This increases the effectiveness of education on hate speech recognition.
When a user searches the web, it brings text. The text enters Azure's virtual machine server and determines whether it is a hate speech in a trained Python model. If the sentence is determined to be "hate speech", masking the sentence according to the option selected by the user. Filtered text/video is shown on the user's web browser.
The deep learning model we're trying to design is a simple task defined by a binary classification(hate or not hate), and had to ensure minimal response speed and stability in order to avoid compromising user experience. Therefore, rather than using heavy SOTA models such as Bert,
We tried to design a neural net structure that could maintain a relatively light level of scale and computational complexity. As a result, we adopted the Bi-LSTM model because the Bi-LSTM model showed the highest accuracy and moderate processing speed. Using the Bi-LSTM Model, a hate speech judgment is made for some sentences.
For sentences that are judged to be hate speech, We used the Lime(Local Interpretable Model-Agnostic Explanation) technique to extract sound masking time and sentences that hate speech filtered. The Lime technique enables precise analysis of the evaluated sentences. Through the Lime technique, we can see why the sentence was judged as a hate speech and how much some words contributed to the result. Using this, we were able to identify the word portion to filter in sentence, and extract the time to filter from the video.
- Express Framework
- Azure VM
https://techcommunity.microsoft.com/t5/student-developer-blog/2021-imagine-cup-category-feature-lifestyle/ba-p/2200906
https://news.joins.com/article/24017473
https://www.donga.com/news/article/all/20210322/106017354/1