Using Latent Topic Models to Examine Teacher Learning in an Online Course

• In Event: 63.085 - MTCC Roundtable Session 26
  In Roundtable Session: 63.085-1 - Investigating Opportunities for Teacher Learning in Online Video-Based Labs

Abstract

Across educational research, there has been increasing interest in Learning Analytics in which techniques from machine learning and data mining are applied to the analysis of education-related data (Authors, 2013; Baker & Yacef, 2009; Siemens & Long, 2011). This interest has been spurred by the increasing availability of data generated as learners work in online environments.

In this study, I apply techniques from Learning Analytics to explore the content of posts made by teachers within the Learning Labs. Specifically, I use latent topic models to examine how teachers understood “models” and “modeling” across the various enactments of our Labs. More broadly, I seek to examine the viability of using computational techniques to examine teacher learning in online professional development.

Methods
The raw data used for this analysis consisted of all of the posts made by participants in our Learning Labs. I used a particular type of latent topic model, latent Dirichlet allocation (LDA), a technique for discovering a set of latent or hidden topics that generate a text (Blei, Ng, & Jordan, 2003). To apply LDA to teacher understanding of models and modeling, my algorithm first identified every use of a form of the word “model” (model, models, modeling, etc.), and then extracted a region of text around the word. This resulted in a collection of ~1400 pseudo-documents. These pseudo-documents were then used as the input to LDA. I varied the parameters used in the model in order to discover a good modeling of the pseudo-documents in terms of a set of latent topics.

Results
Not surprisingly, teacher understandings of models and modeling differed significantly across the two topic areas, science and mathematics. To a lesser extent, these understandings also differed across activities within the same Lab. In mathematics, for example, teacher understanding of models and modeling reflected latent topics that focused on real world and making sense. They also reflected latent topics associated with the particular form that the modeling activities took within the math Lab, namely counting. In contrast, in science, discussions of models and modeling reflected latent topics associated with scientific practices, and the production of drawings of scientific phenomena.

These results suggest that this approach is promising—that LDA can produce analyses that are both interpretable and sensible. I do not believe that automated topic modeling will replace manual analyses of the data. However, automated and manual analyses, when paired, can enrich our understanding and confidence in more traditional types of analysis. The automated techniques also have the potential to allow us to analyze much larger corpora of online data.

Author

• Bruce Sherin, Northwestern University