How can I reproduce results from https://www.aclweb.org/anthology/W19-5358.pdf ,especially (2.2 Using BERT for lexical unit semantic similarity)? I have found only w2v. Could you add examples of configs/embeddings from latest release?
We'd be interested in using YiSi (for example in our analysis toolkit compare-mt), and it'd be very convenient to do so if there was a way to access it through Python. It'd be particularly nice if this could be done by installing through pip. I know this might be a little bit of work, but I just thought I'd make the request in case.
when trying to make Yisi work, I had to fiddle around with different Java versions. In the process I noticed that the make clean target does not clean up everything, which was very confusing because it left Java binaries in tact that I had compiled before switching to a different Java version.
The things that would need to be removed in addition to what make clean already does are mostly the "Build/install directories" mentioned in the .gitignore. I blieve that users are right to expect make clean to clean those up as well.
the reference output files in ./test/ref/ contain absolute file system paths that the test suite does not properly ignore when comparing the test output to them. This makes tests fail on almost all computers.
As a workaround one can simply modify the reference files to contain the "right" paths. But that's of course a bad solution, because it takes time to figure out and actually do. The better way would be to make YiSi output relative paths. Another way would be to make the test script ignore paths.