Closed gavinhenderson closed 10 months ago
Here’s what I think. Roadmap the next release. Test harness first then tweaks. We should get feedback on now as it is. It may impact the roadmap.
Here's a thing we should do. A quick one. Get 10 year sentences. Some long. Some short. Some should have words not in the prediction list eg names. If we implement learning prediction - Do it on a fresh install - twice. Once without and then with learnt words.
Just note time & hits in a one switch scan to say each 10 sentences. Put it in a nice table for each release or / and each prediction system we build
Notes from chat:
Note: Keith suggested using hugging face
Closing in favour of #33 #34 #40
Currently the prediction algorithm is something I just made up. It is 'fast' but it is a very naive implementation of prediction.
Before any work is done on improving prediction i think we need to create a basic test harness that can evaluate any given prediction otherwise we will just be shooting in the dark. The kind of thing I have in mind is we run a few sentences through the prediction engine and count how many clicks it would take to get to the target sentence. Its not a perfect metric but it will at least tell us if we are going in the right direction.
For ref will wrote up these notes: #2
@willwade Open question, whats your thoughts regarding if this should block release or not? Do we push this to a future version or do we improve the prediction ahead of release?