[English] Wrong POS for possessive adjectives

[Polirecyliente]

Good day team,

I've been learning to use spaCy and it's really great. One doubt I had is about the accuracy of the English model en_core_web_lg and other sizes (e.g. en_core_web_sm). I wrote this Python script to display that possible inaccuracy:

import spacy

nlp = spacy.load('en_core_web_lg')
text = 'my house is mine'
doc = nlp(text)

print(doc[0].pos_, doc[0].tag_, doc[0].dep_)
print(doc[3].pos_, doc[3].tag_, doc[3].dep_)

This is the output of those print statements:

PRON PRP$ poss
ADJ JJ acomp

So .. the word "my" is being treated as a pronoun, when in reality it's an adjective (or a possessive determiner if you will). You can verify this in the Merriam Webster dictionary: https://www.merriam-webster.com/dictionary/my and also in the grammar for possessive pronouns from the British Council: https://learnenglish.britishcouncil.org/free-resources/grammar/english-grammar-reference/possessives-pronouns

I think maybe the reason that "my" is treated as a pronoun, might be due to tradition, but a tradition can also become wrong, if the grammatical definitions change over time. Or perhaps is there another reason for this result from spaCy?

It also happens that in the sentence "my house is mine", the word "mine" is treated as an adjective!! That looks to me kind of like worst case scenario. The word "my" which should be treated as an adjective of "house" (or a determiner of "house") is treated as a pronoun ... while "mine" which is the possessive pronoun, is treated as the adjective, so they seem as if they were inverted. ... or if you tell me that "mine" is an adjective in this case, because of the copular verb "is", if that's the case then both "my" and "mine" would be adjectives here.

Please help me know if I'm in the wrong here with counterexamples, or maybe with some citation or resources that explain this kind of situation, because I must admit, I might be completely out of base here. Thank you for your attention.

[Polirecyliente]

Ok so I've been doing my homework, I've been reading more about this, and I can accept that there are ways to properly consider "my" to be a pronoun, because "my" can be seen as a proper pronoun, due to the fact that it is an inflection of "I" which is the pronoun per se.

Now that being said, is there any place in the spaCy data model, or somewhere, that the adjective side of "my" is shown or can be queried or matched against with patterns?

For example take a look at this beauty that you get using the same sentence "my house is mine" using Stanza:

[
    {
      "id": 1,
      "text": "my",
      "lemma": "my",
      "upos": "PRON",
      "xpos": "PRP$",
      "feats": "Case=Gen|Number=Sing|Person=1|Poss=Yes|PronType=Prs",
      "head": 2,
      "deprel": "nmod:poss",
      "start_char": 0,
      "end_char": 2,
      "ner": "O",
      "multi_ner": [
        "O"
      ]
    }
]

As you can see, Stanza also treats "my" as a pronoun, which fair enough, because of it being an inflection of "I" the pronoun itself. But the interesting part is where it says "deprel": "nmod:poss", in this dependency value, Stanza is saying that "my" is not only a 'poss' (which is also said by spaCy), but it also is saying that "my" is an 'nmod', a noun modifier (in this case "my" modifies "house"), so it has functions of an adjective. Is there some way for spaCy to reveal this same type of 'nmod' information about "my"? Thank you 🙏

PD: According to Universal Dependencies at https://universaldependencies.org/u/dep/ there are two types of 'poss': 'nmod:poss' and 'det:poss', so it would seem that spaCy should use one of these? In Universal Dependencies I can't seem to be able to find 'poss' alone on its own. When seeing the dependency relations in English https://universaldependencies.org/en/dep/ the only option that appears is 'nmod:poss'.

[apoorva-01]

The my tag is expected, mine is the actual bug.

For my, PRON PRP$ poss is by design. spaCy's English models use the Penn Treebank tag set, where PRP$ is a possessive pronoun, and the coarse pos_ maps that to PRON. The "determiner/adjective" side you're looking for isn't thrown away, it's just stored somewhere other than pos_. Look at the dependency and the morph:

import spacy
nlp = spacy.load("en_core_web_lg")
t = nlp("my house is mine")[0]
print(t.dep_)    # poss   -> it's modifying "house" as a possessive
print(t.morph)   # Poss=Yes|PronType=Prs

So if what you care about is "is this word acting as a possessive determiner on a noun", check token.dep_ == "poss" and token.morph.get("Poss"), not pos_. That's the grammatical role you were after. pos_ is deliberately coarse.

mine is a different story, that one's a genuine model error. It should be a PRON (predicative possessive, "the house is mine"), and instead your lg model calls it ADJ JJ. I reproduced the same class of miss on en_core_web_sm, where it comes out NOUN NN. Predicative possessives like this are just hard for the statistical tagger and it slips. Nothing to configure, it's an accuracy limit of the model. If it matters for your pipeline, a short attribute_ruler rule keyed on the lemma/text (mine, yours, hers, ours, theirs in a ... is ___ frame) is the usual patch.