Bug: The different punctuation at the end of a sentense lead analysis results wrong.

[qingyun1988]

How to reproduce the behaviour

There is a sentence "Whose bike is broken?". However, the analysis result is wrong.
If I added a Chinese question mark "？" to the end of the sentence, the analysis result would be correct!
By multiple tests, I found that If I added different punctuation to the end of the sentence the analysis results would be different.
These different punctuations are "?", "？", "." and none.
The details are in the attached files. Please have a look at it.
How can I resolve the bug? thank you!

Your Environment

• Operating System: windows 10
• Python Version Used: v3.7.8
• spaCy Version Used: v3.0 and the model version is en_core_web_trf 3.0
• Environment Information:

[polm]

First, note that the results of the small/medium/large models match the output you list as 3, so this is specific to the transformers model.

Second, note that errors on specific individual sentences are inevitable given the way statistical models work, so it doesn't make sense to consider this a bug. You can read more about that in #3052.

Third, I don't think either of these parses is wrong, though the Transformers one is better.

The parse with the passive construction is like if the sentence means "whose bike is being broken?". In that sense the original sentence is grammatical English, but it's not normal, it's like Biblical speech or something.

The Transformers parse is more normal, compare it to the parse for "The book is blue".

https://explosion.ai/demos/displacy?text=The%20book%20is%20blue&model=en_core_web_sm&cpu=0&cph=0

[qingyun1988]

@

[polm]

So on examination, I didn't notice the subj/attr difference. Using attr instead of subj for a noun before the verb is kind of weird. On the other hand, even so it's better than the passive construction.

[qingyun1988]

Thank you for your reply.

I think this sentence "Whose bike is broken?" is very normal and simple. So, it isn't a particular and individual one.

However, the analysis result in the "en_core_web_trf v3.0" model lacks "subj".

The "trf" model is considered more accurate, but this output on this sentence is out of expectation.

What's more, why these analysis results are varied is only because of different punctuation attached to the end of the same sentence. It seems very unstable and weird.

So I consider it a bug.

The example "The book is blue" you taken is correct in performing on "https://explosion.ai/demos/displacy". This is because using "sm" model but not "trf" one. It just proves that there is a problem in "trf" model or in Spacy system.