in this video about twee, David Read demonstrates the process of creating materials about “pub culture in the UK”.

The steps are quite sensible from the point of view of materials development:

  1. begin with a topic, and you can select a level of proficiency in order to grade the language.
  2. list some items of target vocabulary connected with the topic
  3. the “AI” generates text of the appropriate level where the target vocabulary can be highlighted.
  4. go on to produce practice exercises and tasks that use the same target vocabulary.

My observations on this:

  • the general workflow in twee corresponds well to the steps you would follow as a materials writer and certainly speeds up the process.
  • the language model doesn’t actually understand the data it is manipulating. Therefore you can get some mistakes or confusing ideas like the example phrase (from the materials in the video):

I prefer to drink lager rather than beer

  • the above kind of artifacts of the algorithm suggest that you would still need a person with a good understanding of English and the topic in order to proof-read everything. In the above example we could change the sentence to:

I prefer to drink lager rather than ale
We might add the note that the word “bitter” has fallen into disuse. Wikipedia describes bitter as an English style of pale ale