A view months ago I posted on the use of avatars in learning environments.
Since then we have worked out the idea and written a paper on it. Mauro Cherubini neatly summarised the idea in the paper on his weblog:
The premise of this paper is that learners cannot be expected to oversee the whole of their communication and also that chat communication tends to be less structured than face-to-face communication (Stromso et al., 2007). Therefore they aim to build a real-time feedback system that can regulate the collaborative interactions. This workshop paper presents a nice approach to use a part-of-speech tagger and a Bayesian classifier to categorize chat messages into 4 functional categories: regulatory, domain specific, social and technical messages. The authors used manual coders to assign each message to a category. Then they used this corpus to train the Bayesian classifier, showing high accuracy results.
The body parts of the avatars corresponding to the two learners communicating with each other grow or shrink when a new chat message arrives. For instance, when a domain oriented message ("the speed increases") is typed the head becomes a little bigger and other body parts become a little smaller, or when a regulatory message is typed ("I agree, the answer is 4") the body becomes a little bigger. Watching the shape of the avatars change when new messages come in is great fun, possibly even for the learners themselves. The reviewers did not think the avatars to be an appropriate visualisation of learner behaviour. They suggested to use histograms :(. I'm not entirely sure, but when learning the laws of momentum, the last thing a learner may want to look at is a histogram.
In order to make the avatars change shape two methods to analyse the chats were used: looking at the words, and looking at the grammatical structure (part-of-speech or POS-tagging) of a chat message. Both methods classify the chats well, looking at the words produces slightly better results, possibly because the vocabulary of the learning environment is very small and all domain-oriented words (speed, momentum, increases, etc.) get assigned to the domain class (the head of the avatar). One of the strongest grammatical structures the automatic analysis picks up for regulatory message is a verb followed by a personal pronoun. Funnily enough, this structure does not exist in English ("think I" is not grammatical). In our Dutch chats grammatical structure is more or less sufficient to select the regulatory chats, and the underlying algorithms can discover this automatically. The paper contains all technical details.
This is a small contribution to the emerging field of Educational Data Mining. Personally, I think it is stimulating the application of automated analysis techniques has the potential of improving both the understanding of learning environments and makes them nicer to use (especially when we find an even better visualisation of the avatars).
Reference: Anjo Anjewierden, Bas Kollöffel, and Casper Hulshof. Towards educational data mining: Using data mining methods for automated chat analysis to understand and support inquiry learning processes. In Proceedings of International Workshop on Applying Data Mining in e-Learning (ADML 2007) as part of the 2nd European Conference on Technology Enhanced Learning (EC-TEL 2007), Crete, Greece, 2007 (September).