Italian negative pronouns tagged PronType=Ind instead of PronType=Neg

[fingoldo]

Hi, we stumbled on this inconsistency while auditing Stanza's morphological tagging quality across languages.

Summary

Across all Italian UD treebanks, the pronouns nulla, niente, and nessuno are almost
universally tagged PronType=Ind. Every other UD language tags their exact equivalents as
PronType=Neg. I believe this is an annotation inconsistency rather than a deliberate
convention, and propose harmonizing to PronType=Neg.

The UD definition

From the UD feature documentation:

Neg: negative pronoun or determiner. In some languages, weights between negative and
indefinite are very fuzzy. [...] Negative pronouns are used to express the non-existence
of something.

This describes nulla, niente, and nessuno exactly. They inherently denote
non-existence — "nothing", "nobody" — without requiring an additional negation particle.
Compare with genuinely indefinite pronouns like qualcosa (something), qualcuno
(someone), chiunque (whoever), where PronType=Ind is appropriate.

Cross-linguistic evidence

I checked the dev/train splits of the major UD treebanks for each language. Every language
except Italian uses PronType=Neg for the same lexical class:

Language	Treebank	"nothing"	"nobody"	PronType	Count
Russian	SynTagRus	ничто (ничего)	никто	Neg	168 PRON, 100% Neg
Spanish	AnCora	nada	nadie	Neg	34 PRON, 100% Neg
French	GSD	rien	personne	Neg	5 PRON, 100% Neg
German	GSD	nichts	niemand	Neg	12 PRON, 100% Neg
Italian	ISDT/VIT/ParTUT/PoSTWITA/TWITTIRO	nulla/niente	nessuno	Ind	~110 PRON, 99% Ind

Zero variation in RU/ES/FR/DE — every single instance is PronType=Neg. Italian is the
sole outlier.

Detailed Italian treebank breakdown

nulla (113 occurrences across 5 treebanks)

Treebank	PRON (Ind)	ADV	NOUN	Notes
ISDT	36	0	2	NOUN = "il nulla" (the nothingness), "nulla osta" (permit) — legitimate
VIT	27	5	1	2 ADV tokens have PronType=Neg; 3 ADV have no PronType
ParTUT	7	0	0
PoSTWITA	30	0	1
TWITTIRO	4	0	0

The NOUN cases are legitimate (substantivized "the nothingness"). The ADV cases in VIT are
a mix: one genuine adverbial use ("per nulla comparabile" = "not at all comparable"), the
rest appear to be mistagged pronominal objects (e.g., "Nulla dice la normativa" = fronted
obj, not adverb).

Only 3 tokens total (2 nulla + 1 niente, all VIT) use PronType=Neg — suggesting
someone recognized the issue but the fix wasn't applied systematically.

niente (145 occurrences)

Treebank	PRON (Ind)	DET (Ind)	ADV	NOUN	Notes
ISDT	19	1	1	0
VIT	25	4	17	6	NOUN cases look like PRON errors; ADV mixed
ParTUT	2	0	0	0
PoSTWITA	41	17	1	0	DET = "niente politici" (no politicians) — genuine
TWITTIRO	8	1	1	0

The DET usage ("niente + noun" = "no X") is a genuine syntactic function, but even there,
the semantics are negative ("no politicians"), not indefinite ("some politicians").

Linguistic argument

Nulla/niente/nessuno carry inherent negation. "Nessuno ha detto nulla" doesn't need
non before nulla for it to mean "nothing" — the pronoun itself negates. This is the
defining property of PronType=Neg per UD.

PronType=Ind describes a different class. Indefinite pronouns refer to an unspecified
entity: qualcosa (something), qualcuno (someone), uno (one). Nulla doesn't refer to
an unspecified entity — it denotes the absence of any entity. That's negation, not
indefiniteness.

Italian negative pronouns behave identically to their cross-linguistic equivalents:

Italian nessuno = Spanish nadie = French personne = German niemand = Russian никто

Same lexical class, same semantics, same syntactic distribution. The only difference is the
PronType label.

Downstream impact. Any NLP task relying on PronType for negation detection, sentiment
analysis, or scope resolution gets the wrong signal from Ind. Neg correctly marks these
as negation-bearing elements.

Proposal

Change PronType=Ind → PronType=Neg on nulla, niente, and nessuno (and their
inflected forms like nessuna, nessun) when functioning as pronouns (UPOS=PRON) across
all Italian treebanks. This would bring Italian in line with the UD feature definition and
with every other language's practice.

The DET uses of niente/nessuno ("nessun problema", "niente politici") could also be
updated to PronType=Neg, since the determiner is semantically negative ("no problem", not
"some problem"), but I'd defer to the Italian treebank maintainers on that.

Happy to help with the actual data changes if this proposal is accepted.

[Stormur]

The value PronType=Neg is ill-defined in UD and in contrast with other aspects of the guidelines, in particular the presence of Polarity=Neg..

That's negation, not indefiniteness.

The point is that these are not exclusive features.

What you see in Italian data might appear as an inconsistency, but it should actually be the way to go: the best annotation for elements such as it. nulla should be PronType=Ind and Polarity=Neg. There are two things going on at once here: by putting the Neg value into two different features, the overview is lost. This negative elements are very often in relation with a basic, non-negative counterpart. And this gets worse when you consider other negated PronTypes, such as lat. neuter 'neither', which is negatively contrasting, and is not indefinite. There are also negative relative PronTypes, e.g. necubi 'lest (any/no)where', and theoretically many other types, too. If PronType=Neg is meant for all these types, it is obliterating information; if it is just meant as the combination Polarity=Neg|PronType=Ind, there is no reason not to use this less cumulative one.

Just for information, this "split annotation" already happens in (some) Latin treebanks (which were not in your sample), and probably in others too. What the Italian treebanks are missing is Polarity=Neg.

---

The NOUN cases are legitimate (substantivized "the nothingness").

They actually are not, they should always be annotated as PRON while taking a determiner.

one genuine adverbial use ("per nulla comparabile" = "not at all comparable")

This is an oblique ("adverbial", as it were) argument, so it should not get ADV.

---

PS: did you use AI for this post?

I cannot find these passages:

From the UD feature documentation:

Neg: negative pronoun or determiner. In some languages, weights between negative and
indefinite are very fuzzy. [...] Negative pronouns are used to express the non-existence
of something.

[AngledLuffa]

Thanks for the fast reply!

What the Italian treebanks are missing is Polarity=Neg.

Is this something we can/should add to the Italian treebanks for nulla?

The NOUN cases are legitimate (substantivized "the nothingness").

They actually are not, they should always be annotated as PRON while taking a determiner.

Does this need an edit to PRON, then?

one genuine adverbial use ("per nulla comparabile" = "not at all comparable")

This is an oblique ("adverbial", as it were) argument, so it should not get ADV.

Is there an action item here as well, then, to update that ADV to PRON?

I also observed that the other features were not consistent across treebanks - Gender & Number in particular. I added some issues myself to the various treebanks inquiring about that already.

PS: did you use AI for this post?

In my experience, Google likes to summarize the UD documentation for us...

[fingoldo]

PS: did you use AI for this post?

Yeah, asked it to format it nicely in markdown & to rewrite our findings in a more convincing fashion (

[Stormur]

Thanks for the fast reply!

My pleasure!

What the Italian treebanks are missing is Polarity=Neg.

Is this something we can/should add to the Italian treebanks for nulla?

Well, if you ask me, yes! I admit not everybody would agree. But it is also true that PronType=Neg vs Polarity=Neg has already puzzled annotators and I think it has already raised some discussion.

The NOUN cases are legitimate (substantivized "the nothingness").

They actually are not, they should always be annotated as PRON while taking a determiner.

Does this need an edit to PRON, then?

If you ask me, again, yes! There is not much sense in having context-changing part-of-speech annotations, as is also explicitly discussed in one of the basic papers presenting UD.

one genuine adverbial use ("per nulla comparabile" = "not at all comparable")

This is an oblique ("adverbial", as it were) argument, so it should not get ADV.

Is there an action item here as well, then, to update that ADV to PRON?

I do not get exactly what you mean with "action item", but, again, yes! It is always the same word: it can get an article, be in a nominal context with a preposition, etc.

[AngledLuffa]

action item

A thing we need to do!

With some of these treebanks, it's not clear how much the original authors are still maintaining them or whether the right thing is to edit the conllu directly, or if there's a different representation that gets converted to conllu later.

Of course, ideally we'd get agreement across the group of which standard to adopt for nulla

Anyone know what the status is of the Italian treebanks? Last time I interacted with the authors on github was a couple years ago

@msimi
@alessandrateresacignarella
@msang