Does the “joint model” mean joint probability?

[msklvsk]

Does UDPipe-Future predict morphology first, and than use that morphology as an input to syntax or does it consider joint probabilities of such syntax over such morpho? There should be cases of probable morphology which leads to improbable syntax.

[foxik]

In the shared task we tried two approaches, one was to do morphology first and then use it as input to syntax. The other possibility was to compute shared representations for both morphology and syntax and then predict both independently. Generally both approaches seem to result in very similar performance.

[msklvsk]

While investigating morpho errors of UDPipe 1.2 for Ukrainian, I found that ~20% are cases where the tagger selected more popular interpretation instead of rare but correct one. If it considered the unlikeliness of a parse such popular interp will produce, it may have made the correct decision. But it must be different with UDPipe-Future. Looking forward to test it!