DEVELOPMENT OF FACEBOOK SCRAPER AND HATE SPEECH CLASSIFICATION OF SOCIAL MEDIA USERS USING PRE-TRAINED MODEL FEATURES IN INSTITUT TEKNOLOGI BANDUNG’S SOCIAL NETWORK ANALYSIS APPLICATION

• Main Author: Junod, Girvin
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/74152
• Institution: Institut Teknologi Bandung

Before the development of this Final Project, the Social Network Analysis Application by Institut Teknologi Bandung (SNA ITB) was only able to procure data for analyzation from the social media Twitter. This is not ideal as the number of users for other social medias like Facebook dwarfs the number of Twitter users in Indonesia. Because of that, the Facebook scraper feature was added to the SNA ITB application in this final project. Another feature was also developed for the application which was the hate speech classification feature. The significant amount of hate speech that is spreading around social media and its negative impact on society causes the feature to be highly sought and it was decided that the feature would be added to the SNA ITB application. The development of the Facebook scraper feature is done using tools that can scrape data from Facebook based on keywords or user profiles. Several APIs and web pages related to the uses of Facebook scraper in the application is created in this Final Project so that the feature can be used. The addition of Facebook scraper causes a change in the application flow for several features. Because of that, several changes are also done to existing web pages and APIs to fit the new application flow. The addition of the hate speech classification feature is done by developing a hate speech classification model using a pre-trained model and developing various APIs and web pages related to hate speech statistics. Other than that, an experiment is also attempted to try to improve the performance the hate speech classification by injecting knowledge into the pre-trained model. The development of the hate speech classification model is successful with the final model demonstrating a strong performance with an accuracy score of 73.89% and F1 score of 80.77% for multi-label classification. This final model is a pre-trained IndoBERTweet model that is fine-tuned using an Indonesian hate speech dataset. The attempt to improve the model’s performance using knowledge injection failed to produce a better performing model. The result of this Final Project is the SNA ITB Application that now has new features such as the Facebook scraper feature and the hate speech classification feature. Users of the SNA ITB application now can scrape data from facebook and analyze it using the features of the application. Hate speech analysis can also be done to the scraped data.