Unable to reproduce SPACE DASH SPACE example from demo

[knuser]

Hi,

Thanks for wonderful work. I'm trying to reproduce demo steps in GUI and stuck on example about rewriting dash rules: https://youtu.be/Dil5i0AYyu8?t=188

As you can see only one example is found and no spaces are added:

Is this a bug or am I doing something wrong?

[knuser]

Also during installation of requirements.txt in fresh virtual env I see a version conflict: ERROR: spacy 2.1.3 has requirement jsonschema<3.0.0,>=2.6.0, but you'll have jsonschema 3.1.1 which is incompatible.

After reinstalling spacy (pip install spacy==2.1.3) to fix jsonschema version conflict and installing tutorials/requirements_tutorial.txt i see again conflict:

ERROR: jupyterlab-server 0.3.4 has requirement jsonschema>=3.0.1, but you'll have jsonschema 2.6.0 which is incompatible.

[tongshuangwu]

As you can see only one example is found and no spaces are added:

Thanks for reporting this! I quickly checked, and turned out it's a bug I introduced during refactoring. Could you please pull from the master branch, and see if it's working now?

ERROR: spacy 2.1.3 has requirement jsonschema<3.0.0,>=2.6.0, but you'll have jsonschema 3.1.1 which is incompatible.

This is a known problem with some of the dependencies having conflicts with each other, but like you have seen, it doesn't break other things, so it's safe to leave it there!

[knuser]

Could you please pull from the master branch, and see if it's working now?

It is working now on incorrect_preprocessing group. Thank you.

I tried also applying dash separation on all_instances, after fresh server start (no other data groups or rewrite rules were present) and saw that hint below text is sometimes incorrect:

[tongshuangwu]

Sorry for my late reply! I was overwhelmed by some other projects. Could you please elaborate what's the issue? From the figure, I didn't see anything abnormal?

[knuser]

Hi, sure:

[tongshuangwu]

I see. I'm guessing your question is, "why the inline edit does not match with the blue predictions?" Sorry if it's confusing, but this is not a bug :)

In the command, what we are requesting is to replace "-" with " - " in the sentence that contains the groundtruth -- not the groundtruth span. Like in the first example:

• The groundtruth (and the prediction, judging by the blue color indicating correct prediction) is "South Pacific"
• "South Pacific" is in the first sentence of the context
• In this sentence (not the groundtruth span), there is one "-" in 1883-84
• Errudite edits it, hence the inline edit status shows "~1883-84~" -> "1883 - 84".
• The span "South Pacific" is not affected by the edit, and after the edit, the model prediction is still this span, hence the prediction delta status shows "South Pacific" -> "South Pacific".

The core idea is, editing the groundtruth sentence does not necessarily affect the model prediction, and these two examples in the screenshot happen to be un-affected.

In incorrect_preprocessing, we required the prediction span to be close to the groundtruth span (hence usually in the same sentence) and to have "-". In other words, whenever you modify "-" -> " - ", you are modifying the prediction span, and hence the inline edit can be matched with the blue/orange prediction status.

Does this make sense?

[knuser]

Oh, now I get it. Thanks for clear explanation, it makes perfect sense.

Thank you