BERT-Based Detection of Sexual Harassment in Dialogues

It tends to become a trend of booking transportation through network equipment with the further integration of the Internet into people's life, and online car-hailing platforms have sprung up. However, a new social crisis has also come with it. Because of the need for platform expansion, most online ride-hailing drivers have not undergone strict professional ethics reviews, which increases the risk of passengers taking the car. Primarily, female users are more susceptible to abuse and harassment and even persecution by drivers due to their disadvantaged position. Unfortunately, this phenomenon is happening every day and even getting worse. Regarding this aspect of supervision, it is difficult for relevant departments to have a more direct management plan. However, it is difficult for relevant departments to give a more natural and effective management plan. Therefore, ensuring the safety of passengers (predominantly female passengers) using online car-hailing becomes particularly important. In the Chinese field, few people try to improve this problem from the perspective of natural language. This work expects to use natural language technology to evaluate the driver's language and determine the degree of potential danger and criminal tendency, thus protecting the passenger and providing evidence for the judicial authorities. We first collected many dialogues between drivers and passengers, then used back translation to expand the corpus. Finally, we adopted various BERT-based model methods to compare and analyze the performance of different variants.