Development of Classification Method for Lecturer Area of Expertise Based on Scientific Publication Using BERT

Abstract

Implementing the Artificial Intelligence concept in higher education can be utilized in the context of Human Resource (HR) talent management. The lecturer portfolio provided by the Integrated Resource Information System (SISTER DIKTI) is expected to give an overview of the profiles of all lecturers and map competencies based on groups of fields of knowledge. However, the map of scientific fields based on SISTER data currently available is still subjective. The data is in the form of a group of lecturers' chosen fields of science, independently selected by each lecturer to recognize their expertise. This study discusses the problem of processing unstructured SISTER data. It looks for mapping solutions and classification methods by developing a strategy for classifying groups of scientific fields from unstructured data input. It is necessary to identify the suitability of the chosen field of science compared to that developed through the tri-dharma through identification based on a mapping of the group of fields that can be extracted from the tri-dharma activity, in this case, research represented by scientific publications recorded on SISTER. Therefore, we need an appropriate model to measure similarity, which can then be classified based on abstract documents and scientific publication titles for the classification of scientific fields using NLP based on classification run on DGX A100. This study aims to develop a classification method from titles and abstracts. Scientific publications contained in SISTER are unstructured data, so a corpus is formed to identify the lecturer's field of science. The results show that the classification method developed in this study can measure the similarity of lecturer publications based on abstracts and titles through a vector formation process based on bidirectional encoders and also produces a deep learning model to classify 24 categories of fields of science with an accuracy of 95,0345 percent on training data and 92.876 percent on the test data.