Wen, Yun

Online discourse is widely used in diverse contexts of learning and professional training, but superficial interactions and digression often occur. In the face of these problems and the large-scale unstructured text data, the traditional way of learning analytics has been challenged in terms of providing timely intervention and feedback. In this paper, a workflow for automatically detecting in-service teachers’ cognitive engagement in an online professional learning community is described. Discourse data of 1834 in-service teachers involved in a teacher professional development program was collected and processed using the Word2vec toolkit to generate lexical vectors. The method of vector space projection was used to calculate the new information contained in each post, cosine similarity was used to calculate topic relevance, and cluster analysis was used to explore in-service teachers’ discourse characteristics. Results showed that in-service teachers’ average contribution was 4.59 posts and the average length of each post was 39.47 characters in Chinese. In the mathematics online professional learning community, the average amount of new information contained in each post was 0.221 and in-service teachers’ posts contained much new information in the early stages of online discourse. Most in-service teachers’ posts were relevant to the discussion topic. Cluster analysis showed three different groups of posts with unique characteristics: high topic relevance with much new information, high topic relevance with little new information, and low topic relevance with little new information. Finally, limitations are discussed and future research directions are proposed.