Customized Named Entity Recognition Using Bert for the Social Learning Management System Platform CourseNetworking

Abstract

Named Entity Recognition (NER) is an information extraction task and one of the most researched applications to extract knowledge from massive data. Conventional NER systems identify predefined entities like name, person, location, organization, time, etc. However, there is a limitation to identifying user-defined entities that are specific to an application. This challenge introduces the concept of customized NER. For instance, if a learning management system like CourseNetworking (CN) needs to identify the skill set of a user from their posts, the existing pre-trained NER models cannot be used. To overcome this information extraction limitation, we propose a customized named entity recognition system for the CN platform using the deep learning model, Bi-Directional Encoder Representation from Transformer (BERT) which is a transformer-based deep learning technique where all output elements are connected to all input elements with dynamic weight connections. The proposed customization model can be employed to train any entity of user choice with a decent amount of training dataset. The model shows 70-72% recall and F1-Score varied on the number of epochs trained. This model is used in various applications like fraudulent detection, recommendation systems, and business intelligence.