Automatic entity recognition based on BERT in computer supported collaborative learning

Abstract

Knowledge building plays a critical role in promoting knowledge acquisition and facilitating the retention of target knowledge in computer supported collaborative learning (CSCL). The interactive texts in CSCL environment provide a valuable opportunity for instructors to understand and evaluate the knowledge building process and results. Entity recognition for interactive texts is the first vital step in evaluating the level of knowledge building. However, the methods of manual recognition and key-term matching are widely applied, which not only time consuming and lack semantic understanding for interactive texts, but also the accuracy of recognition is hardly guaranteed. We proposed an automatic, accurate combination method to recognize knowledge entity based on a state-of-the-art natural language processing model-BERT (Bidirectional encoder representation from transformers) to understand semantic meaning of interactive texts in CSCL. Text classification and entity recognition are employed in this study. Adopting BERT automatically classify the whole interactive texts into knowledge and non-knowledge types. Levenshtein Distance (LD) and semantic matching based BERT are used to recognize entity from literal and semantic similarity between student interactive texts and entity corpus provided by teachers. Using 16047 interactive texts produced by 51 groups of college students around the strategies of problemsolving in educational psychology are analyzed. The classification accuracy is 90.07%. 7025 knowledge interactive texts were used to automatic entity recognition and F1 value of concept entity and principle entity recognition are 72.02% and 61.18% respectively, while processes entity is 48.75% and examples entity is 44.32%. The automatic combination method shows potential value in assisting teachers in understanding the level of knowledge building and provide feedback timely in CSCL context.