Feature-based comparison in pyclts: allow for the use of "foreign" feature systems

[LinguList]

6000 of PanPhon's segments have been linked to pyclts, and they include a feature system. But how can one use the features when trying to compute distances between sounds? pyclts should offer some ways of accessing different feature systems and computing feature similarities. Even if these are explicitly done for one feature system, it would be an interesting integrative service, since we have quite a few feature systems now:

• phoible
• panphon
• nidaba
• mielke

[XachaB]

I would have a use for this feature !

I have some implementation for this for a project of mine. I parse each of the feature systems, then the distances can be either of hamming distance, Jaccard index over features; or the more elaborate Jaccard index over natural classes defined by the features, as described by Frisch (1996), using the concepts python library. This last similarity metric is really nice, but works best with monovalent feature systems.

If there is interest, I could prepare a PR. I would need some pointers for how to do this in the CLTS eco-system (where would you want to integrate this functionnality ?), rather than as an external library.

[LinguList]

Thanks, this is interesting, just read again your study, where you present the usage of the Jaccard. Unfortunately, I am completely undecided how to add this now. We would need a regular format to store feature values and feature systems, integrated in CSVW or CLDF, and then we should be able to parse it. So we'd need an initial discussion of how a feature system might look in a first instance, i.e., how differently they might look.

[XachaB]

Ok, I can imagine how that requires a lot more design decisions. I'll subscribe to the issue in case I can contribute something later on.