Xinrui Que¹ and Yao Pan², ¹USA, ²Vanderbilt University, USA

Community based websites such as social networks and online forums usually require users to register by providing profile information and avatars. It is important to ensure these user uploaded information comply with the website policy. This includes the information being personal, related and clear, as well as not containing unhealthy/disturbing content. A review or censorship system is usually deployed to review new user registration. Nowadays, many platforms still use manual review or rely on 3rd party APIs. However, manual review is time consuming and costly. While 3rd party services are not tailored to the specific business needs thus do not provide enough accuracy.

In this paper, we developed an automatically new user registration review system with deep learning. We apply the state-of-art techniques such as CNN and BERT for an end-to-end evaluation system for multi-modal content. We tested our system in E-pal, a freelancing platform for gaming companionship and conducted a qualitative evaluation of the approach. The results show that our system can evaluate the quality of avatars, voice descriptions, and text profiles with high accuracy. The system can significantly reduce the effort of manual review and also provides input for the recommendation ranking.

Deep learning, Image classification, BERT, CNN.