French morphologizer mislabeling future, conditional, imperative

[KennethBaclawskiJrDialpad]

Hello,

I'm using spaCy to model French conversations, and I see that the morphologizer is not performing as well as I'd expect for unambiguous irrealis forms (specifically future tense, conditional mood, imperatives). I understand the underlying reason is probably that these forms aren't frequent in the training data, but are there any potential updates or recommended workarounds?

For example:

• "Remplacez le verbe par un nom.": The imperative "Remplacez" is misrecognized as a noun (using POS/TAG). This is similar to [French morphologizer] Mislabelisation of Mood=Imp|Number=Sing|Tense=Present #8147, but broader in that "Remplacez" is unambiguously a verb.
• "Vous remplacerez le verbe par un nom.": The future tense "remplacerez" is misrecognized as an imperative (MORPH contains Mood=Imp, Tense=Pres), even though it is unambiguously the future.
• "Vous remplaceriez le verbe par un nom.": The conditional mood "remplaceriez" is misrecognized as the future tense (MORPH contains Mood=Ind, Tense=Fut), even though it is unambiguously the conditional.

I'm working with a set of ~160 common French verbs and tested their whole paradigms in this way. 98% of infinitives are recognized correctly, but only 13% of second person plural imperatives (34% even had incorrect POS like in the example above), 37% of future tense, and 7% of conditional mood forms. Sure enough, I see that these three categories are uncommon in the UD French Sequoia data.

How to reproduce the behaviour

import spacy
nlp = spacy.load("fr_core_news_sm")
doc = nlp("Remplacez le verbe par un nom.")

for token in doc:
    print(token.text, token.pos_, token.morph)

Info about spaCy

spaCy version: 3.8.2
Python version: 3.11.9
Pipelines: fr_core_news_sm (3.8.0)

[thjbdvlt]

You could try the pipeline i've trained, which is available here: https://github.com/thjbdvlt/solipCysme

I've made it because I have the same issue you. It's trained mostly on novels (19e-21e) and texts with a lot of interactions, personnal pronouns and differents moods. The morphologizer uses HunSpell output (po: and is:) as features (along with PREFIX and NORM), so it should have good results even on texts very different that the ones my model has been trained on.

It has some limits. First, there is no Gender morphological feature, so it would have terrible scores on (e. g.) UD datasets. It's not absolutely accurate on NOUN/ADJ for sentences like les jeunes bourgeois (where both words could be ADJ/NOUN). And some other stuff (like ADJ/Passive Past Participe; or "que" as PRON and "que" as SCONJ).

Of course I could provide some information if you need it. Let me know!