What’s in a Cluster? Automatically Detecting Interesting Interactions in Student E-Discussions

Abstract

Abstract. Students in classrooms are starting to use visual argumentation tools for e-discussions -a form of debate in which contributions are written into graphical shapes and linked to one another according to whether they, for instance, support or oppose one another. In order to moderate several simultaneous e-discussions effectively, teachers must be alerted regarding events of interest. We focused on the identification of clusters of contributions representing interaction patterns that are of pedagogical interest (e.g., a student clarifies his or her opinion and then gets feedback from other students). We designed an algorithm that takes an example cluster as input and uses inexact graph matching, text analysis, and machine learning classifiers to search for similar patterns in a given corpus. The method was evaluated on an annotated dataset of real ediscussions and was able to detect almost 80% of the annotated clusters while providing acceptable precision performance.

Exploring creative thinking in graphically mediated synchronous dialogues

“…Interestingly, the results we achieved with this data improved upon most of the results reported on a different data set, with different annotated cluster types (Mikšátko & McLaren, 2008). More specifically, the DOCE algorithm generally performed better in finding snippets of ''creative reasoning" than it did in finding more standard argumentation structures, such as ''chain of opposition" (i.e., a chain of contributions by students in which they go back in forth in argue for and against a given issue) and ''argument + evaluation" (i.e., a student makes an argument which is then evaluated by another student).…”

“…However, in an earlier experiment, reported in (Mikšátko & McLaren, 2008), we compared DOCE to a simple program that returned random clusters and found that DOCE performed significantly better. While the random algorithm is, admittedly, a ''low bar" to exceed, doing significantly better than random demonstrated that DOCE is clearly finding (at least some) clusters of interest.…”

“…Or, it can be used as a ''live" classifier of clusterscharacteristic example(s) representing a cluster of a particular type are stored in the database and used later as queries for automated cluster detection. Details about the underlying DOCE algorithm are provided in (Mikšátko & McLaren, 2008).…”

“…Example (DOCE) (Mikšátko & McLaren, 2008) that allows us to identify places in e-discussions in which students may be critically deepening or creatively widening the conversation, as well as other types of complex conversational moves. DOCE is one of a number of tools that have been developed on the ARGUNAUT project to assist a teacher in monitoring the on-going simultaneous e-discussions of several groups of collaborating students.…”

“…Applying a comprehensive coding scheme aimed at identifying structural argumentative and dialogical events in the synchronous discussions mostly does this. These expertly coded events were then used to develop classifiers using artificial intelligence techniques Mikšátko & McLaren, 2008;Scheuer & McLaren, 2008) that are able to detect and classify these events automatically and inform the moderator. This is done by a component of the ARGUNAUT system called the Deep Loop.…”