French lemmatizer/Pos tagger issues, lemmatizer suggestion

[databill86]

trafficstars

Hello,

As a follow up to #11298, I would like to share some examples for which the french tagger and lemmatizer fail to return the correct results (spaCy v3.4.1 and fr_core_news_lg v3.4.0).

I'm using the "rule" mode for the french lemmatizer, but it fails to give the right lemma for some "simple" examples (eg., je vous confirme mon numéro, it returns "confirme" instead of "confirmer". So I tried to figure out how to solve this by debugging its behavior on rule_lemmatize, and I came across this:

https://github.com/explosion/spaCy/blob/3e4cf1bbe1745a55ede0dece31353aebc3f82729/spacy/lang/fr/lemmatizer.py#L74-L77

If we don't find a rule for the token we check in the lookup table. However, while debugging, I think that string in lookup_table.keys() will never be satisfied:

• string = token.text, and
• lookup_table.keys() is a list of odict_keys odict_keys([11901859001352538922, 11991225512756778597,...)

If I'm not mistaken, I think that's what's meant to be tested is: token.norm in lookup_table.keys().

If that's the case, I think we can safely refactor the code as:

        if not forms:
            forms.extend(oov_forms)        
            forms.append(self.lookup_lemmatize(token)[0])
        

[UPDATE]: after some more checking, I found that string in lookup_table.keys() will invoke __getitem__ of Table. But I still don't get why don't we simply use self.lookup_lemmatize(token)[0] as I suggested.

Even if the token is not in the lookup table, self.lookup_lemmatize(token)[0] will do the same thing in line 77.

This will solve the lemmatization problems for examples like: "je vous confirme mon numéro..." (confirme will be lemmatized to "confirmer" instead of "confirme").

Here are some other examples for which we have tagging and lemmatization issues:

Text	Word analysed	Lemma detected	Lemma expected	Pos detected	Pos expected	Comments
je remarque qu'elle n'est toujours pas parti....	parti	parti	partir	ADJ	VERB	"Parti" is a past participle it should be analyzed as a verb and not as an adjective
alors que je l'avais annulé via le formulaire sur Internet.	Internet	Internet	internet	PROPN	NOUN
Je vous confirme que mon numéro de téléphone est bien le [phone_number].	confirme	confirme	confirmer	VERB	VERB	Problem"pos tagging" + LEMMA => same problem with the VERB "informe"
Comme vous me l’avez demandé par mail je vous communique mon numéro	communique	communique	communiquer	VERB	VERB	Problem"pos tagging" + LEMMA
Je vous redonne mon numéro	redonne	redonne	redonner	VERB	VERB	Problem"pos tagging" + LEMMA
le numéro de téléphone de votre service client communiqué en bas de vos mails	communiqué	communiquer	communiquer	ADJ	VERB	Problem"pos tagging"
Avez-vous un service client téléphone?	téléphone	téléphon	téléphone	ADJ	NOUN	Problem"pos tagging" + LEMMA
MERCI DE SUPPRIMER DEFINITIVEMENT MON COMPTE ET TOUTES LES DONNEES PERSONNELLES ASSOCIEES A CELUI-CI.	SUPPRIMER	supprimer	supprimer	NOUN	VERB	problem "pos tag"
MERCI DE SUPPRIMER DEFINITIVEMENT MON COMPTE ET TOUTES LES DONNEES PERSONNELLES ASSOCIEES A CELUI-CI.	DEFINITIVEMENT	definitivemer	définitivement	VERB	ADV	problem "pos tag" + lemma
MERCI DE SUPPRIMER DEFINITIVEMENT MON COMPTE ET TOUTES LES DONNEES PERSONNELLES ASSOCIEES A CELUI-CI.	DONNEES	donnee	donnée	NOUN	NOUN	problem "lemma" + lemma
MERCI DE SUPPRIMER DEFINITIVEMENT MON COMPTE ET TOUTES LES DONNEES PERSONNELLES ASSOCIEES A CELUI-CI.	ASSOCIEES	associee	associer	NOUN	VERB	problem "pos tag" + lemma
JE SOUHAITE IMPERATIVEMENT ETRE SUPPRIMEE DE VOS LISTES DE DIFFUSION.	SOUHAITE	souhait	souhaiter	ADJ	VERB	problem "pos tag" + lemma
JE SOUHAITE IMPERATIVEMENT ETRE SUPPRIMEE DE VOS LISTES DE DIFFUSION.	IMPERATIVEMENT	imperativement	impérativement	NOUN	ADV	problem "pos tag" + lemma
JE SOUHAITE IMPERATIVEMENT ETRE SUPPRIMEE DE VOS LISTES DE DIFFUSION.	SUPPRIMEE	supprime	supprimer	ADJ	VERB	problem "pos tag" + lemma
JE SOUHAITE IMPERATIVEMENT ETRE SUPPRIMEE DE VOS LISTES DE DIFFUSION.	DIFFUSION	DIFFUSION	diffusion	PROPN	NOUN	problem "pos tag" + lemma

How to reproduce the behaviour

Your Environment

• Operating System:
• Python Version Used: 3.8
• spaCy Version Used: 3.4

[polm]

Thanks for pointing this out, this does look like a bug in the lemmatizer, and seems to also affect Catalan. I've opened a PR with a fix at #11382.

About the refactoring - having a separate fallback in the French lemmatizer itself may be to avoid relying on the implementation of the lookup tables in case they're swapped out or something, so we need to check that.

[polm]

Regarding the POS tag issues, I haven't looked at that in detail yet, but keeping in mind #3052, is this a degradation from a previous version, or is it plausible this is just normal errors?

[databill86]

Regarding the POS tag issues, I haven't looked at that in detail yet, but keeping in mind #3052, is this a degradation from a previous version, or is it plausible this is just normal errors?

It's actually a degradation from the previous version, and I actually didn't find what exactly changed between v3.2 and v3.4.

[polm]

Thanks for the clarification, let us know if you have more details.

To clarify the status of this issue, the fix for the lemmatizer issue has been merged, so that should be resolved, and we just need to look into the POS tag issue.

[rmitsch]

Marking this as resolved. Feel free to open a new issue if the POS problems persist!

[github-actions[bot]]

This issue has been automatically closed because it was answered and there was no follow-up discussion.

[github-actions[bot]]

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.