G is for Granularity

3 05 2015

Granular is a buzz word in the discourse of publishing these days. With its vaguely breakfast cereal connotations it conjures up an image of learning content made palatable and wholesome.

For example, Knewton, the company that specializes in adaptive learning software, features a short video clip on its website, in which the presenter advises us that

“Publishers need to be looking at producing granular content. … no longer in the form of a big-package textbook, but broken down into small chunks that teachers, students, administrators can choose to use in combination or in a blend with any other content that they choose to use”.

Grains – chunks – blends: it’s making my mouth water.

Elsewhere on the Knewton site, we get this heady, but somewhat less appetizing stuff:

Within the adaptive learning industry, a shared infrastructure can benefit all existing educational apps by providing them with unlimited back-end content, granular and highly accurate student proficiency data, robust analytics, and more.

And

Differentiated learning can help each student maximize their potential by shaping the curriculum so that each student understands their proficiencies at a granular level and is given a direct path to improving them.

In a recent blog, they even show us what the granules (aka taxons) of second language acquisition look like:

Knewton taxons

Click to enlarge

But there are at least four major flaws in the way language learning has been granularized. These flaws long pre-date data analytics, but by bringing the power of industrial-scale computing to bear on data collection and analysis, companies like Knewton (and the publishers who enlist their services) are magnifying these flaws exponentially.

The first flaw – let’s call it the taxon fallacy – is that they have got their granules wrong. Notice that the so-called taxons in the Knewton graphic are the traditional ‘tenses and conjugations’ (present continuous, past perfect etc) – the same ‘tenses and conjugations’ that have been passed on like a bad gene from one generation to the next ever since the dawn of recorded time (or ever since the teaching of Latin) but which have little or nothing to do with how the English language is either used or internalized.

The units of language acquisition are not ‘tenses and conjugations’ (English has no conjugations, for a start). The units of language acquisition are words and constructions. Construction is a general term for any form-meaning association — whether a single word, a phrase, or a more abstract pattern — that has become conventionalized by the speakers of a language (see this related post).  Constructions are more than just ‘lexical chunks’ – they can also include morpheme combinations (e.g. verb + -ing) and syntactic patterns (e.g. verbs with two objects) – and they are much, much more than ‘tenses and conjugations’. They are not easily located in the syllabus of a standard coursebook – the type of syllabus which is still the default setting for data analysts such as Knewton.

The second fallacy – I’ll call it the proceduralization fallacy – is another legacy of a long tradition of transmissive teaching: it is the belief that declarative knowledge (e.g. knowing that the past of ‘go’ is ‘went’) automatically converts to procedural knowledge, i.e. that it is available for use in real-time communication. Hence, the assumption is that, if the learner is tested on their knowledge of an item (or granule) and found to know it, it follows that they will be able to use it. As teachers we know this is nonsense. Researchers concur: Schmidt’s (1983: 172) long-term case study of a Japanese speaker of English led him to conclude that ‘grammatical competence derived through formal training is not a good predictor of communicative skills.’ Counting the granules tells you very little about a learner’s communicative capacity.

Related to this fallacy is what is known as the accumulated entities fallacy, described by Rutherford (1988: 4) as the view that ‘language learning … entails the successive mastery of steadily accumulating structural entities, and language teaching brings the entities to the learner’s attention’. Since at least the 1980s we have known that, as Ellis (2008: 863) puts it, ‘grammar instruction may prove powerless to alter the natural sequence of acquisition of developmental structures.’ And Diane Larsen-Freeman (1997: 151), coming from a dynamic systems perspective, reminds us that

Learning linguistic items is not a linear process – learners do not master one item and then move on to another. In fact, the learning curve for a single item is not linear either. The curve is filled with peaks and valleys, progress and backslidings.

Unless a granular approach to data collection and analysis factors in these ‘peaks and valleys’, it will have nothing very interesting to say about a learner’s progress.

Finally, there is the homogenization fallacy: the view that all learners are the same, have the same needs, and follow the same learning trajectory to the same ultimate goals. This quaint belief explains why the designers of adaptive learning software think that it is possible to calibrate any single learner’s diet of granules on the basis of how 50,000, or indeed 50 million, other learners consumed their granules. Although software designers using data analytics pay lip-service to ‘differentiation’ and ‘personalization’, essentially they have a battery chicken view of language learning, i.e. that the same grains are good for everyone, even if they are meted out in slightly different quantities and at slightly different rates.

Contrast that view with the sociolinguistic one that no two people speak the ‘same language’: ‘You and I may both be speakers of language X but your grammar and mine at the descriptive level will not be identical … We both appeal to different sets of rules’ (Davies 1991: 40). Or, as Blommaert (2010: 103) writes, ‘Our real “language” is very much a biographical given, the structure of which reflects our own histories and those of the communities in which we spent our lives.’ It does not exist in someone else’s data-base, much less in granular form.

In the end, as Brumfit (1979: 190) memorably put it, ‘language teaching is not packaged for learners, it is made by them. Language is whole people’.

Ergo, it is not granular.

References

Blommaert, J. (2010) The Sociolinguistics of Globalization. Cambridge: Cambridge University Press.

Brumfit, C. (1979) ‘Communicative’ language teaching: an educational perspective. In Brumfit C.J, and Johnson, K. (eds.) The Communicative Approach to Language Teaching. Oxford: Oxford University Press.

Davies, A. (1991) The Native Speaker in Applied Linguistics. Edinburgh: Edinburgh University Press.

Ellis, R. (2008) The Study of Second Language Acquisition (2nd edn). Oxford: Oxford University Press.

Larsen-Freeman, D. (1997) ‘Chaos/Complexity science and second language acquisition’. Applied Linguistics 18/1.

Rutherford, W. (1988) Second Language Grammar: Learning and Teaching. London: Longman.

Schmidt, R. (1983) ‘Interaction, acculturation and the acquisition of communicative competence,’ in Wolfson, N., & Judd, E. (eds.) Sociolinguistics and Second Language Acquisition, Rowley, MA: Newbury House.

Photos taken from http://flickr.com/eltpics by Hada Litim, used under a CC Attribution Non-Commercial license, http://creativecommons.org/licenses/by-nc/3.0/

Note: Coincidentally, Philip Kerr has just blogged on this same topic, i.e. Knewton’s ‘Content insights’, here: Adaptive Learning in ELT