[IC] Tests to determine whether the "Regroup" function is working

[JoyfulGen]

@martha-thomae this is what you were asking about!

So a couple of months ago I was beginning to wonder whether the "Regroup" part of "Regroup and reclassify" was actually working, because the proportion of neumes actually getting regrouped seemed very low. I made a synthetic neume layer to test the grouping function:

I had squares and I had rectangles, and I also had shapes made of two rectangles placed close together. In the IC, I did the following:

1. Classified one rectangle as "rectangle" and one square as "square"
2. Regroup and reclassify ; All of the two-rectangle shapes got classified as "rectangle"
3. I then classified a two-rectangle shape as a "square"
4. Regroup and reclassify ; The remaining two-rectangle shapes still got classified as "rectangle."

So that test didn't work. I made a new test image that was more fool-proof, to make sure of my results:

This time, instead of putting rectangles side by side to create a new shape, I took a square and put a thin white line through it. This way, the split shape would have exactly the same dimensions as the squares. Then I did this:

1. Classified one rectangle as "rectangle" and one square as "square"
2. Regroup and reclassify ; All of the split-square shapes got classified as "rectangle"
3. I then classified a split-square shape as a "square"
4. Regroup and reclassify ; The remaining split-square shapes got classified as "square."

So that test was a success! It would appear that the "Regroup" function does work, but it's very very sensitive and only works under the best conditions. It's also possible that changing the regrouping parameters might help; in both my tests I just used all of the default parameters. @martha-thomae let me know what you think and I'll be happy to perform more tests if that can be useful!

[martha-thomae]

Thank you, Gen! This is really useful, thanks for sharing :)

[martha-thomae]

So, now, knowing that

It would appear that the "Regroup" function does work, but it's very very sensitive and only works under the best conditions.

What would you suggest for the following problem / situation?

The Situation:

I have two symbols in Aquitanian notation:

• Torculus, a few examples below with different intervals
  

• Cephalicus
  

As you can see, the last part of the torculus looks pretty similar to a cephalicus. When I classify torculuses (torculi?), I usually need to group the two parts before assigning them the class torculus. When I do regroup and reclassify, rarely the torculi get "regrouped" and classified correctly. But if I leave them separated into their two parts (the first punctum and the following two notes), it might be easier for the IC to confuse the last part with a cephalicus... I think. I haven't conducted strict tests on all of this. But I was wondering, in your experience, what would be best: to try to always group the torculus and classify them as such? or to leave them as two parts (class punctum and class second_half_torculus) and later expect for the MEI Encoding to group these two parts into one <neume>? Always considering that is also very important that the IC somehow distinguishes the second_half_torculus from a class cephalicus.

I hope I was clear enough. It is hard for me to explain the issue, I just want for both signs to be handled as correctly and efficiently as possible during the path from IC to MEI Encoding to Neon correction.

[JoyfulGen]

Oof, that's a tricky one! It's good that you mention that it's important that the IC differentiate between the second_half_torculus and the cephalicus, because otherwise I would tell you to teach the IC to call those two symbols the same thing and sort it out later. But I understand that won't work for your notation, so another solution is needed!

This is tricky because I'm not sure about either option! I admit, I don't feel confident that the IC will be able to group the two halves of a torculus consistently. If we use my test as reference, the two halves are way too far apart for the IC to understand that they should be grouped together. So my hunch would be to go with your second option of leaving the torculus as two parts and let MEI_encoding group them later. I completely agree with you that the IC might confuse the second_half_torculus with the cephalicus, but the IC has sometimes surprised me with its capacity to distinguish symbols, especially if I was very strict in my training. If you always choose nice, crisp examples of both glyphs and make sure to keep an equal number of instances of both in your training data, you'll increase your chances of the IC successfully telling the difference between the two.

So for what it's worth, I would suggest trying leaving the torculus as two parts. If you have the time for it, it could be valuable to test this with a few folios, to get a feel of what the IC thinks... I'm sorry I don't have a better solution for you, I hope this is still helpful! Let me know how it goes. I'd also be happy to test something for you if that can be helpful!