The downside of `ExtPos` in the features

[AngledLuffa]

So, I've found what I think is a meaningful downside to having ExtPos in the features instead of the Misc column. It's really difficult for a model to learn this in imbalanced situations. So for example, a in Spanish occurs 13,000 times in AnCora, 1000 times in an MWT. This is actually not an unreasonable ratio to learn. But then de occurs about 40,000 times, only 400 of which are MWT. Not surprisingly, Stanza's models completely punt on this.

@dan-zeman - I remember you saying you didn't like ExtPos in the features either...

[nschneid]

MWT: do you mean a fixed expression?

Is this an issue of Stanza not being able to predict infrequent fixed expressions? Perhaps it needs a lexicon-based feature for matching known expressions?

[AngledLuffa]

That would certainly be an option... hadn't even considered that, but it would solve the problem I'm running into. Thanks for the suggestion!

Nevertheless I can't help but think it shows a bit of a design flaw that it's 1% of de words with this feature in Spanish, for example

... and yes, I did mean fixed expression rather than MWT, thanks for the clarification

[nschneid]

Typo=Yes is another sparse feature. I get that they're a bit different from most of the morphosyntactic features, but the information has to go somewhere, and better if there is an incentive for parsers to predict it. :)

[dan-zeman]

I remember you saying you didn't like ExtPos in the features either...

Yes, I also wanted it to live in MISC and not in FEATS. Not so much because it is rare and unpredictable—I think we still want to be able to predict it, and I suspect that was motivation for other people to smuggle it in FEATS and thus in parsers' way (although if that was the motivation, we could easily dismiss it, as people can train a model that will predict things in MISC as well). But I don't like that it is not a feature of the word it appears on. Typically it is a feature of a multiword expression, and even if it is a POS changer of a single word, it completely depends on context and not the word itself.

[sylvainkahane]

Many words are polycategorial. For instance, "love" can be VERB and NOUN and only the context can help you to decide which upos is the good one. In the case of "love", it is clear, we have two different upos. But some other cases are less clear it is useful to have the possibility to "play" with two POS labels.

In English-GUM the most common PROPN is "University" and you also have "States", "Science", "Research", "Academy", "Army", etc. I would really prefer to have these nouns analyzed as NOUNs and have an additional ExtPos=PROPN indicating that they are (part of) a proper name in this context.

In spoken productions, we have unfinished words. Most of the time we can guess the POS (and even the lemma). But we don't want these forms to be in the lexicon and we decided to give them upos=X and to add an ExtPos for the POS given by the context.

Another problem is discourse markers. Most of them are interjections, but some come from other POS. Due to this very particular role, they tend to lexicalize with a new sense and it is not very clear which POS we must choose, the original one or INTJ, considering the word is now polycategorial. See examples of the different POS of discourse markers in English. For the spoken French treebanks, we solved this dilemma by keeping the original POS in upos and adding en ExtPos=InTJ.

Symbols can have a very clear POS. For instance, $ in 10$, or % in 10% are clear nouns. Due to the UD constraint to have upos=SYM, an ExtPos=NOUN is very welcome.

Here, you have many other examples of ExtPos used for individual words in the French GSD treebank.

[amir-zeldes]

In English-GUM the most common PROPN is "University" and you also have "States", "Science", "Research", "Academy", "Army", etc. I would really prefer to have these nouns analyzed as NOUNs

To be clear, under the current English guidelines at least, PROPN already implies that the word is a noun. Parts of names that are not nouns do have the xpos tag NNP (or NNPS in plural), but they have the expected upos tag based on the universal guidelines. For example:

1	For	for	ADP	IN	_	4	case	4:case	Discourse=joint-sequence_m:104->98:2:lex-indph-954-955-_|PDTB=Implicit:Temporal.Asynchronous.Precedence:then:_:921-942:943-981
2	the	the	DET	DT	Definite=Def|PronType=Art	4	det	4:det	Bridge=177<188:comparison-relative|Entity=(188-event-acc:inf-nnnnn-cf6-3,6,8-coref
3	second	second	ADJ	JJ	Degree=Pos|NumForm=Word|NumType=Ord	4	amod	4:amod	_
4	campaign	campaign	NOUN	NN	Number=Sing	16	obl	16:obl:for	_
5	in	in	ADP	IN	_	10	case	10:case	_
6	the	the	DET	DT	Definite=Def|PronType=Art	10	det	10:det	Entity=(173-abstract-giv:inact-sssnn-cf3-2,4,5-coref
7	Always	Always	ADV	NNP	_	8	advmod	8:advmod	MSeg=Al-way-s|XML=<hi rend:::"italic">
8	Keep	Keep	VERB	NNP	Mood=Imp|Person=2|VerbForm=Fin	10	compound	10:compound	_
9	Fighting	Fighting	VERB	NNP	VerbForm=Ger	8	xcomp	8:xcomp	MSeg=Fight-ing|XML=</hi>
10	series	series	NOUN	NN	Number=Sing	4	nmod	4:nmod:in	Entity=173)
11	in	in	ADP	IN	_	12	case	12:case	_
12	April	April	PROPN	NNP	Number=Sing	4	nmod	4:nmod:in	Entity=(189-time-new-nnnnn-cf10-1-sgl|XML=<date when:::"2015-04">
13	2015	2015	NUM	CD	NumForm=Digit|NumType=Card	12	nmod:unmarked	12:nmod:unmarked	Entity=(175-time-giv:inact-nnnnn-cf5-1-coref)189)188)|SpaceAfter=No|XML=</date>
14	,	,	PUNCT	,	_	4	punct	4:punct	_
15	Padalecki	Padalecki	PROPN	NNP	Number=Sing	16	nsubj	16:nsubj	Entity=(1-person-giv:act-sssss-cf2*-1-coref-Jared_Padalecki)
16	partnered	partner	VERB	VBD	Mood=Ind|Number=Sing|Person=3|Tense=Past|VerbForm=Fin	0	root	0:root	MSeg=partner-ed
17	with	with	ADP	IN	_	18	case	18:case	_
18	co-star	co-star	NOUN	NN	Number=Sing	16	obl	16:obl:with	Entity=(97-person-giv:inact-nnnnn-cf4-1,3-coref-Jensen_Ackles|MSeg=co-star
19	Jensen	Jensen	PROPN	NNP	Number=Sing	18	appos	18:appos	XML=<ref target:::"https://en.wikipedia.org/wiki/Jensen_Ackles">
20	Ackles	Ackles	PROPN	NNP	Number=Sing	19	flat	19:flat	Entity=97)|XML=</ref>

Note how "Padalecki" is PROPN/NNP, but "Fighting" is VERB/NNP.

Generally speaking I agree it's awkward to have tags like SYM which don't really describe a grammatical category (and as you point out, could be a noun or something else grammatically), I think at this point we are stuck with UPOS as it is, so if people want to use ExtPos on top of that I totally support it, I'm just not sure how easy it would be to implement it. But I think tagging parts of names that are not nominals as PROPN is already against the current universal guidelines, and we've made an effort to follow the guidelines in English (though we may have missed some cases!)

[nschneid]

Well https://universaldependencies.org/u/pos/PROPN.html specifies that "Cat" in "Cat on a Hot Tin Roof" is NOUN not PROPN, but we decided that would be too hard to implement for English as it goes against the longstanding PTB convention.

[AngledLuffa]

Tin and Roof presumably also PROPN then? Hot_ADJ? I mean, that's how I annotate things in English treebanks if I see them, but it'd be good to establish that this isn't just cats.

[amir-zeldes]

specifies that "Cat" in "Cat on a Hot Tin Roof" is NOUN not PROPN, but we decided that would be too hard to implement for English

I definitely don't think English is alone in making such an implementation unfeasible, and I've opposed this idea pretty strongly for a number of reasons. If taken seriously, it would lead to quite a lot of results that I think would be very counter-intuitive. All of the following would cease to contain any PROPNs:

• The movie "Terminator", the musical Cats, etc.
• United States, United Arab Emirates, etc.
• King's College etc.

But there are many reasons to believe these are in fact proper names, despite being form-identical to common nouns - we don't say "the King's College", we can say "Cats is a wonderful/terrible musical" (not "are"), etc. And note that if we also treat acronyms as the POS of their head - so "idk" is a VERB - then the following would also become NOUN:

• USA
• UAE
• UCL

etc. None of that seems intuitive to me - there is a difference between "the emirates that I mentioned earlier" and "the Emirates" (meaning the country, UAE). The latter is the name of a country and should be PROPN, the former is just a NOUN (some emirates). I'm pretty sure most UD corpora use PROPN along these lines.