Strange parse of Italian sentence.

[iliakur]

How to reproduce the behaviour

JSON formatting is mine.

>>> import spacy
>>> model = spacy.load("it_core_news_sm", disable=["ner"])
>>> doc = model("la gatta mangia pesce")
>>> doc.print_tree()
[
    {
        "word":"pesce",
        "lemma":"pesce",
        "NE":"",
        "POS_fine":"S__Gender=Masc|Number=Sing",
        "POS_coarse":"NOUN",
        "arc":"ROOT",
        "modifiers":[
            {
                "word":"mangia",
                "lemma":"mangiare",
                "NE":"",
                "POS_fine":"V__Mood=Ind|Number=Sing|Person=3|Tense=Pres|VerbForm=Fin",
                "POS_coarse":"VERB",
                "arc":"nsubj",
                "modifiers":[
                    {
                        "word":"la",
                        "lemma":"la",
                        "NE":"",
                        "POS_fine":"RD__Definite=Def|Gender=Fem|Number=Sing|PronType=Art",
                        "POS_coarse":"DET",
                        "arc":"det",
                        "modifiers":[

                        ]
                    },
                    {
                        "word":"gatta",
                        "lemma":"gatta",
                        "NE":"",
                        "POS_fine":"S__Gender=Fem|Number=Sing",
                        "POS_coarse":"NOUN",
                        "arc":"amod",
                        "modifiers":[

                        ]
                    }
                ]
            }
        ]
    }
]

The parse claims pesce is the ROOT of the sentence, whereas in fact it is the object. The rest of the dependencies are correspondingly unrealistic.
Interesting to note that the POS-tagger seems to work fine, the parser just ignores the postags completely.

Your Environment

• Operating System: Ubuntu 18.04
• Python Version Used: Python 3.6.7
• spaCy Version Used: 2.0.9
• Environment Information:
  models: en, fr, de, nl, it

[honnibal]

The parser and POS tagger are independent --- it doesn't use POS tags as features. So yes, sometimes the parser makes errors that don't match the POS tags, sometimes vice versa.

[iliakur]

We played around a bit more with the model and for what it's worth, these variations of the sentence get parsed correctly:

• la gatta mangia il pesce
• il gatto magia pesce

[ines]

Merging this with #3052. We've now added a master thread for incorrect predictions and related reports – see the issue for more details.

[lock]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.