Fine-Tuning Deep Learning Model for Detecting Indonesian Traffic Accident Tweet

Twitter is one of Indonesia's most used social media, with total visits reaching 192 million visits per month, making it the fifth most-visited website in Indonesia. Twitter users exchange information with each other daily. One of them is a traffic accident event. With over 100,000 cases every year from 2017–2021, it is not usual to see a tweet that contains traffic accident information. By implementing text classification, traffic accidents can be detected through tweets and used to send alert messages to emergency responders to respond as quickly as possible. Previous research used a machine learning approach to solve this text classification problem. The study resulted in the best accuracy of 85% from the SVM model, which leaves room for improvement using state-of-the-art deep learning models. This research proposes a transfer learning method using two BERT-based and two RoBERTa-based pre-trained models, namely IndoBERT, IndoBERTweet, Indonesian RoBERTa Base Wiki, and Indonesian RoBERTa Base OSCAR to create the best model to classify tweets that contain traffic accident information. We used two kinds of fine-tuning architecture: the vanilla version fine-tuning architecture and the modified version, which combines the pre-trained model with CNN. This study used 1,002 manually labeled tweets, where 360 were labeled as traffic accident tweets. The experiment found that transfer learning from a pre-trained model Indonesian RoBERTa OSCAR provided the best result with an f1-score of 88.73 % and an accuracy of 92%, which means a 7% accuracy improvement over the previous research's best model.