Our contribution

We propose that research on digital learning lacks (and desperately needs) precisely the sort of framework provided by work on distributed cognition. While distributed cognition was first applied to socially situated work environments, for example to the study of ship and airplane navigation (Hutchins 1995a and 1995b), recent developments have applied this approach more specifically to HCI (human computer interaction) research (Wright, et al 2000). What has not yet been done, and what we propose, is to apply the notion of distributed cognition to the domain of learning, more specifically, digital learning. This constitutes more than simply extending the reach of existing assumptions in distributed cognition to one more domain, that is, from research on work environments to research on learning environments. Rather, the extension to the field of learning requires a novel framework for evaluating the symbiotic relation between humans and their artifacts that distributed cognition encompasses. For example, one central notion of distributed cognition is 'cognitive off-loading' or the phenomenon of shifting certain cognitive burdens from the human to his or her tools. An electronic calculator with a memory is a perfect example. An architect or engineer benefits from shifting the burden of calculation to the digital tool thus freeing precious cognitive and attentional resources that can then be devoted more productively to other tasks. If the domain, however, is not work but learning, the value of such cognitive off-loading must be reassessed. Does an electronic calculator enhance the learning of a math student? Certainly the framework for evaluating the contribution of cognitive off-loading onto digital tools for a math student must be different than for working architects or engineers. Cognitive off-loading, for example, can obviously hinder learning by allowing a learner to avoid exercising precisely the cognitive skills they are intended to acquire. We suggest that while the notion of cognitive off-loading is playing a crucial role in the growing research on distributed cognition in work environments, to apply the notion effectively to research on digital learning, an alternative set of assumptions and evaluation metric is needed. Part of our current work is devoted to developing this novel framework and evaluation metric for digital language learning. For this purpose we have proposed the notion of learning events, more specifically 'digitally assisted language learning events' to which we turn next.

Event Ontologies and Digitally Assisted Language Learning Events (DALLE)

A precursor to laying out a framework for evaluating digital effects on learning is to have some coherent model of what sorts of events occur in digital environments. Once an ontology of such events is available, these events can be investigated, manipulated, and incorporated explicitly in the design of digital learning environments or tasks. In the case of digital language learning, we have set out to develop an ontology of digitally assisted language learning events. These are discrete types of actions or events that users engage in during the course of unrestricted or restricted activity on line. In the initial stages of this work, we have confined the scope of investigation to web-browsing events online. The learning event ontology will be derived from a set of features deemed relevant to learners' language learning (both incidental and intentional language learning). Examples of feature types include such categories as user purposes in engaging in the event, costs incurred by the event (in terms of the users' time, their attentional resources, etc.) types and density of interactivity (human to human interactivity vs. human to text or image; continuous interactivity vs. intermittent, etc.), locus of event initiation or event propogation (events initiated or propogated by user vs initiated or propogated by system, etc.)

The contribution of such a taxonomy will be to provide a coherent framework for research on digital language learning, for the design or evaluation of digital language learning platforms or activities and for discussion and debate of this research and these platforms. For example, given a particular online scenario, such as a learner accessing a particular webpage via a browser, this learner could have one or more of a range of purposes in this setting. The learners' purpose matters because it will determine what sorts of digital assistance or tools would be helpful and what sorts would be disruptive. If the purpose is to acquire new useful vocabulary, then the costs in time and attentional resources spent on individual unknown words could be a plus since these efforts deepen the learners grasp of the targeted vocabulary. If, however, the learners' main purpose in this same setting is instead to locate specific information or to comprehend the main ideas of the webpage text, then time and attention spent on particular unknown words may be distracting, and tools that highlight or mark these words would be correspondingly disruptive. A more useful tool under these conditions would be one that can detect not simply unknown words, but those words that serve as keywords to the theme of the text. Such knowledge provided to this learner could enhance their ability to reconstruct the main theme of the text.

Thus a taxonomy of learning events based upon a set of features and feature types and incorporating consideration of user purposes among other factors can provide a coherent framework for research on digital language learning and for system design and evaluation. This serves as the motivation for our work in constructing such an taxonomy and the feature types that it will be built upon.

NL₂P

An assumption of our work is that language technology can be developed to play a contributing role in foreign language learning. Such technology will rely upon results from Natural Language Processing. However, direct application of NLP techniques to language learning risks creating a mismatch between what the technology can do and what learners really need. Thus, many of the techniques of NLP must be submitted to careful reconsideration when viewed from the point of view of language learning. Two basic sorts of language data that must be considered are: (1) standard target language that can serve as models or input to language learners, and (2) language produced by learners. Both bring novel considerations to current NLP techniques and approaches. Take for example the language produced by learners. Two areas of application here are the construction of learner corpora and the development of grammar checking to be used on the texts that learners produce. Consider the construction of learner corpora, that is, machine readable collections of text that learners produce in the target language. One of the fundamental tasks in corpus construction is annotation of the texts with linguistic information, such as part-of-speech (POS) tagging. The existing techniques devised for POS tagging standard corpora rely upon entropy models of the language, models of transition probabilities, what POS is most likely to occur in a particular context. Here, the texts produced by learners will not conform to these models. That is, learners way use a noun in a context where such a part of speech is highly unlikely to occur in standard target language use. This feature of learner output wreaks havoc on current POS tagging software when applied to learner output. Moreover, there is the question of whether the text should be tagged with the target grammar's POS or with the POS that the learner's non-target grammar assigns to a word? For example, we have evidence that in a common error type where learners lack the 'be' verb in predicative contexts, as in “They aware I need help,” the learner considers “aware” to be a verb, thus explaining the lack of “be” introducing it. Should the learner corpus tag this instance of “aware” as an adjective, using the target language grammar, or as a verb, using the learners grammar. Is POS tagging of a learner corpus intended as an indication of the learners grasp of the language or not? These questions involve technical challenges for NLP beyond the current state of the art for POS tagging as well as questions concerning how to ascertain the learner's grammar of the target language that accounts for the non-target language uses they produce.

Our goal, then, with respect to NLP is to develop novel well-motivated approaches to concrete challenges presented by language learners and educators.

Ubiquitous Networked Language Learning

Another aspect of Hutchins' work from which our work takes inspiration is his focus on investigating cognition 'in the wild,' that is, in its natural habitat en vivo as opposed to en vitro or under laboratory conditions.