Discovering How Language Patterns Evolve in Online Discourse

• In Event: 72.065 - Roundtable Session 30
  In Roundtable Session: 72.065-7 - Computational Methods for Qualitative Learning Sciences Research: Affording Materiality to Text-Based Data

Abstract

Objectives:
This work describes a new approach for better understanding how online discourse changes as discussants become enculturated. By exploring the evolution of word embeddings in a large online forum, we found that some “trajectories” emerged as forum users became more expert. In the talk, we will detail those trajectories and explore the implications of using word embeddings to understand how language features evolve at the community level.

Theoretical framework:
This work takes the explicit theoretical stance that discourse reflects understanding and enculturation not simply in the explicit vocabulary that participants consciously employ but in the structure of their language, including word substitutions, synonyms, and (likely unconscious) syntax choices. In this way, we align with Discourse analysis investigations into the microgenetics of language (as per Gee, 2014), and the focus of legitimate peripheral participation (Lave & Wenger, 1991) on how people take up community (language) practices.

Methods:
We induce latent word representations (embeddings as per Turian, Ratinov, & Bengio, 2010) that predict markers of writer expertise as a function of linguistic phrase structure. We chart shifts in the latent word space as writers gain experience and expertise with the underlying subject matter. This analysis sheds light on common learning trajectories and the way these trajectories are manifested in abstract syntactic and semantic patterns of language.

Evidence:
The word uses the hundreds of gigabytes of uncompressed text discourse data that have been collected from the online forum boardgamegeek.com. The corpus is very large and represents the entirety of discussion on the public forum.

Results:
The study is ongoing; preliminary work suggests that language transforms in multiple dimensions as a user posts more on the site. Not only do users tend to more complex technical language as they go on, but they describe different targets, goals, and community focus. Rather than potential purchases, they are more likely to use words describing strategy.

Scholarly significance:
The methods being explored have many potential applications across the learning sciences to better understand how discourse changes as people learn using the vast textual sources available on the web. Using large data sets to understand how language shapes the relationship between community, learner, and language has myriad possible benefits, such as improved community design, better understanding of how people take up community language, and how learners manifest legitimate peripheral participation through language features.

Authors

• Matthew W. Berland, University of Wisconsin - Madison

• Benjamin Snyder, University of Wisconsin - Madison