Manifold Learning is not compatible with Annotated Corpus Map

[processo]

I noticed that t-SNE is used in the help example before Annotated Corpus Map to reduce dimensions, PCA and MDS can also be used but not Manifold Learning. Why not allow it?

OS: Windows 10
Orange: 3.37.0

[wvdvegte]

I found out that you can use the output of Manifold Learning as input for Annotated Corpus Map if you put a Corpus widget in between and reassign the 'Used text' features. Apparently, when applied to data containing information on what's considered to be the corpus, that information is removed by Manifold Learning, while it is required in the input of Annotated Corpus Map.

It's a bit confusing, though, that Manifold Learning has t-SNE as one of its Methods to select from, with different options than the ones available in the t-SNE widget. And cosine distance (best for text) is not available as distance metric, although it should be according to the Manifold Learning documentation ...

edit: I reported this as an issue in biolab / orange3

[processo]

Nice find. I did not know Corpus can redefine text attributes at any stage. I guess coordinates are just arbitrary data and Annotated Corpus Map uses whatever the "used features" are on input.

I noticed something else. Manifold removes tokenization too (and probably all text info), so if I don't add a Preprocess after Corpus then punctuation appears in cluster labels. Probably an implicit tokenizer in Corpus or Annotated Corpus Map without any filter.