Kirschbaum (2013)

[LinguList]

Unsupervised Segmentation for Different Types of Morphological Processes Using Multiple Sequence Alignment

Implementation probably challenging, but very interesting, and potentially the most difficult approach to implement for now.

[arubehn]

The implementation of the algorithm is actually quite trivial, and is already there under morseg.segment.kirschenbaum (in the old structure, still subject to adaption).

The hard part here is to cluster together the words that should be aligned to each other. The method makes use of distributional semantics (see Section 3.1 in the paper), which is a piece of information that we naturally cannot use, and of phonetic similarity (Section 3.2), which is effectively measured by some sort of edit distance.

We therefore need a clever method for clustering "semantically and phonologically related words", as the paper calls them, to emulate the method. Once those clusters are established, the calculations are straightforward.

[LinguList]

Semantic relatedness can be handled via Concepticon, since we have linked data to Concepticon. Phonetic similarity can be handle with sound classes as well, which could be faster, depending on what one looks for. So we can tackle this approach as a next starting point. I'd try to adopt it from what you have implemented so far to the style that I propose.