Closed arubehn closed 4 months ago
Results for the distinctiveness analysis are looking completely off now... something must have gone wrong along the recent changes we have made
I fixed the issue regarding the distinctiveness analysis, the numbers are back to the correct order of magnitude. However, there are still some minor deviations that I will investigate.
It shows why extensive tests are so useful.
We should add a list of all features with their names to the test repository and run that also in the future to control every change.
Agreed :)
What I would do - once I have found and fixed the error - is to generate a file containing the ~8,000 CLTS sounds (or a representative subset) with the expected feature vectors. Then we can consistently test against that. Or does reading data from external files mess with the automatic testing workflow?
Yes, that would be cool. This does not take long to test, and we have something for the future!
I have investigated all the mismatching sounds. It appears that they can be classified into three classes - and for two of them the error was actually in the old code, not in the new one:
Since no sound that we have analyzed in the plots or in the concordance lines analyses was affected, these parts can remain as they are. I will rerun the quantitative analysis on distinctiveness; there might be minor changes, but the figures will definitely remain in the same order of magnitude. So, essentially, I think only the numbers have to be changed accordingly :)
Turns out, the numbers (almost) didn't change at all (probably since the bugs only affected relatively marked sounds, and were systematic) - the distinctiveness analysis per language now is back to exactly the same numbers, and from the ~8k CLTS sounds, we are now actually capable of providing one (1!) more unique feature vector (5318 instead of 5317) :D
Well done :-) It makes me also glad to see that we could improve content-wise with the new code.
The two sounds are most likely defined like this by Cormac, and I would trust his judgment here. They are hard-coded into CLTS, not generated, with the double-features (velar-and-uvular) that are generally very rare.
Okay, that‘s perfect - then that case is also solved :)
While rerunning the evaluation scripts to make sure that everything still works smoothly after the refactoring, I found out that the number of unique has slightly changed (from previously 5,317 to 5,255). Seems like the feature mapping for a small group of sounds has been changed accidentally... I will look into it.