Should Number=Ptan be used instead of Number=Plur for English plural-only words?

[rhdunn]

Words such as "glasses" and "species" are in their plural form according to English pluralization rules. Regarding the lemmas, with some exceptions:

1. EWT uses the plural form as the lemma, matching its intended use;
2. GUM uses the singular form as the lemma, presumably using the machine-lemmatized plural form.

EWT is correct here, but these cases are very likely to confuse lemmatizers trained on the UD English corpora due to these cases not following the English plural lemmatization rules.

Making use of the plurale tantum annotation (Number=Ptan) which already exists would make the intention clearer, and allow lemmatizers to differentiate from NNS/Number=Plur and NNS/Number=Ptan.

Note: This should also apply to the dates like 1980s where the lemma retains the form's plural suffix.

Relevant issues:

1. EWT -- Ambiguous lemmatization of pluralia tantum UD_English-EWT#374
2. PUD -- Plural nouns not using singular lemma, possible pluralia tantum UD_English-PUD#33

[nschneid]

I was today years old when I learned of the Ptan value. :) It is documented and several languages do use it, so in principle it would make sense to adopt it for English as well.

It would have the benefit of explaining why lemmas contain the plural morphology; lemmatizers/checkers would have to implement a fixed list of pluralia tantum anyway, and this makes it explicit in the data.

A counterargument might be that it is a rare value that does not really have morphosyntactic consequences for English beyond the lemma—morphosyntactically it is a kind of plural, so users may expect Number=Plur.

[nschneid]

Here is a nice little summary of pluralia tantum and singularia tantum in English: https://english.stackexchange.com/questions/407446/does-english-have-any-singularia-tantum-besides-mass-nouns

@amir-zeldes and I are on board with implementing Number=Ptan for English. It would help if somebody else could submit PRs.

Let's NOT implement Number=Coll because collective nouns comprise a very large set.

[nschneid]

I've rewritten https://universaldependencies.org/en/feat/Number.html to better reflect how we currently use the feature. The page isn't updating immediately, but you can look at the diff.

Note that, contra the original post above, I don't think species falls under the category of pluralia tantum.

Here's how I defined it:

Ptan: plurale tantum

Some nouns appear only in the plural form, with a regular plural suffix and plural agreement, but lack a singular counterpart (at least when serving as a nominal head). (The lemma is therefore the plural form.) These form a relatively closed set. Semantically, they often denote a mass-like collection, or a doublet object.

Note that some nouns have endings that look like regular plural endings, but are not: linguistics and Xerxes are singular, and species and series may be singular or plural, but none of these are pluralia tantum.

Examples

• clothes, scissors, riches

[nschneid]

Note that even for pluralia tantum, the "s" can sometimes be chopped off when used attributively or in a compound ("pant leg", "scissor kick"). Hence the qualification "(at least when serving as a nominal head)".

[amir-zeldes]

Right - looks like I responded in the wrong issue! All I need is a list/notes if you want to kick anything OFF the list in the GUM validator.

[nschneid]

Current GUM validator lists where the lemma is allowed to end with "s":

• NNPS: ["Netherlands","Analytics","Olympics","Commons","Paralympics","Vans", "Andes","Forties","Philippines"]
• NNS: ["surroundings","energetics","politics","jeans","clothes","electronics","means","feces", "biceps","triceps","news","species","economics","arrears","glasses","thanks","series"]
• plus pluralized decades

Additional items in EWT that are being flagged:

• regards troops supplies grounds goods headquarters contents pants manners memoirs respects whereabouts finances proceedings savings specifics genitals remains wares barracks environs orthodontics earnings statistics panties tenterhooks dynamics geopolitics trousers hackles auspices confines fives billiards genetics scissors eatables furnishings

I don't have clear intuitions about all of these, hoping somebody else can weigh in.

Note that only a subset are pluralia tantum—not "economics", "series", "species", or "news" (at least).

[sylvainkahane]

I don't understand what you want to do. In English, there are only two numbers, singular and plural. OK, some nouns can only be singular or only plural. This an interesting property, but which concerns the lexicon. These nouns behave as normal singular or plural nouns. If you add another value for Number, you will break one of the most common rule of English: "The verb agrees in number with its subject." What could be annotated is the fact that a noun is massive or countable. Pluralia tantum are just massive plural nouns, no? I understand the question about the lemma but it must not act on the grammar: "massive vs countable" must be a lexical feature, which play a role in some construction: choice of the determiner, choice of number for generic use, etc. Note also an additional problem: a same lexeme can have different senses which are massive and countable (compare "I like fish" (massive) and "I like fishes" (countable)). Do we want to annotate this? (it is interesting but difficult to do without revising manually the treebanks.)

[AngledLuffa]

OK, some nouns can only be singular or only plural.

That's literally the proposal - annotate those nouns. Ptan stands for "plural tantum" where "tantum" in some way represents the fact that the word is of a class which is never singular. (TBH I'm not really clear where that terminology came from) So for example, you could say "I have chips. Here, have one" but it would be rather weird to say "I have savings. Here, have one"

[rhdunn]

There are various plural-only words such as in "I put on my glasses". With these, the lemma is the same as the form, not the depluralized form. I.e. the lemma for "glasses" in this case is "glasses" not "glass". The proposal is to mark these uses as Number=Ptan instead of Number=Plur.

If Number=Plur is used, it is harder for NLP systems to learn the depluralization rules, resulting in inconsistent lemmas in generated output. Especially if there are plural instances.

[rhdunn]

These are Latin terms:

1. plurale tantum -- plural only (https://en.wikipedia.org/wiki/Plurale_tantum, https://en.wiktionary.org/wiki/plurale_tantum#English)
2. singulare tantum -- singular only (https://en.wiktionary.org/wiki/singulare_tantum#English)

[sylvainkahane]

We know what "plurale tantum" means. It is not the question. A plurale tantum is a lexical unit whose occurences are all plural. The properties of the lexical unit must not be confused with those of its occurrences. Number=Plur is a property of the occurences. Number=Ptan does not make any sense for me (at least in English). This property is a property attached to the lexical unit itself. It must be encoded by another feature.

[nschneid]

@sylvainkahane I understand Number=Ptan to be a subtype of Number=Plur. As you say, it is a subtype that concerns itself with the overall lexeme as reflected in the lemma. Another way to do this would be something like LexNumber=PlurOnly|Number=Plur. I don't have any particular preference between the two implementations but Number=Ptan is already part of the universal guidelines. Perhaps @dan-zeman would like to weigh in.

[rhdunn]

Would Number=Plur,Ptan also work? -- UD allows multi-valued features.

[nschneid]

I do have a slight worry that algorithms projecting agreement features onto verbs from their subjects would naïvely copy Ptan, when it really only applies to nouns. So that would be a practical argument in favor of separating LexNumber.

[amir-zeldes]

Ptan is a canonical annotation value of UD, and I assume in any language where it is used, it implies Plur, so @sylvainkahane 's objection is not really English-specific IMO, it sounds like a general criticism of Ptan. But other languages do use it in exactly this way, so it is UD English that's unusual here.

I think it's not so odd to have more specific values that imply other values. For example, in many languages, numbers are essentially nouns (e.g. Semitic), or there is no really strong distinction between ADJ and NOUN, but we still use the more specific tag where appropriate. Should we avoid NUM in Arabic just because it obscures the facts that Arabic cardinal numbers are morphosyntactically also nouns? I think in such cases it can be understood that a language should implement the most specific labels possible, and an implication hierarchy such as Ptan -> (subtype of) Plur is understood.

I do have a slight worry that algorithms projecting agreement features onto verbs from their subject

Such alogrithms wouldn't get very far anyway: even just for plain coordination of two singulars you need to switch to Plur, so I would say the problem would be with the algorithm, not the annotation. I don't mind if UD says as a whole that Ptan is not a value of Number, but if it is, then I see nothing about the English case to suggest this shouldn't be used here.

[dan-zeman]

Yes, Number=Ptan can be viewed as a special case of Number=Plur, so if a language uses it, then agreement between Ptan subject and Plur verb is not unexpected.

In some languages (English, probably), the special bit is a property of the lexeme, as @sylvainkahane points out. In others (Czech, for instance) it also has morphosyntactic implications (you must use different forms of numerals with plurale tantum than with normal plural nouns).

[amir-zeldes]

OK, so if we're doing this we need a list. Here's what I gleaned from the above plus the GUM exempt plural form lemmas (presumably EWT has some more):

lemma	xpos	notes
Netherlands	NNPS
Analytics	NNPS
Olympics	NNPS
Commons	NNPS
Paralympics	NNPS
Vans	NNPS
Andes	NNPS
Philippines	NNPS
Maldives	NNPS
surroundings	NNS
energetics	NNS
politics	NNS
jeans	NNS
clothes	NNS
electronics	NNS
means	NNS
feces	NNS
remains	NNS
news	NNS
economics	NNS	listed by @nschneid as non Ptan, but if functioning as NNS I think it is
arrears	NNS
glasses	NNS
thanks	NNS
ergonomics	NNS
aesthetics	NNS
twenties	NNS	and 20s, 1920s etc. for years (but not: as plural for sets of 20)
thirties	NNS
…	NNS
pants	NNS
scissors	NNS

Not Ptan: species, series, biceps, triceps

Any additions/comments welcome!

[AngledLuffa]

Disagree on series: one TV show or one set of 7 playoff games is one series

[amir-zeldes]

Disagree on series: one TV show or one set of 7 playoff games is one series

Right, one series - so there can be a single one, or multiple series. So it's not Ptan, just a noun whose singular form is identical to the plural form (like "sheep"), no?

[nschneid]

Yeah: "That series is canceled." "Those series are canceled." It ends in -ies because of the Latin source, nothing to do with pluralization.

Ptan means the form cannot be used in the singular or made singular in the same sense. (Maybe "economics" is valid as Ptan in the plural: If I say "the economics are sound" that's has nothing to do with multiple "economic"s, and is a different sense from talking about the field of economics.)

[AngledLuffa]

My mistake, sounds good

[nschneid]

Implemented for EWT. Scripts: UniversalDependencies/UD_English-EWT@547b675...cd0d92f#diff-e02db0ba7788687b383704df1414689e399e8d52709bac84029ab9a86d64c109

I disambiguated "respects" manually, but haven't worked on the "-ics" nouns ("politics", etc.), except to apply Ptan to the ones in the list. In EWT the default for such nouns is NNS whereas in GUM it is NN. Absent evidence from subject-verb agreement it can be hard to tell whether an "-ics" token should be considered singular or plural....

[nschneid]

Number=Ptan was implemented for English, at least EWT and GUM!