Verbnouns

[jheinecke]

In Welsh, the base forms of verbs are verbnouns, which functions as nouns. There are no infinitives. Whereas infinitives (like English or German) mark direct objects in the same way for finite and infinite verbforms, verbnouns do not, cf. German

• ich sehe ihn (I see him[ACC])
• ihn[ACC] sehen (to see him)

and Welsh

• Mi welais i hi (lit. Affirmative see-PAST-1SG she, "I saw her")
• ei gweld (lit. possessive-3SG-FEM see-VERBNOUN, "seeing her/to see her")

If the auxiliary bod (be) is used in its verbnoun form, we get a noun under a cop-deprel:
bod yn goch "being red" (yn is a predication particle, needed if an adjective or noun is the predicate)

1  bod  bod  NOUN  verbnoun  _  3  cop
2  yn   yn   PART  pred      _  3  case:pred
3  goch  coch  ADJ   adj     _  0  root

As the issue 3, this is rejected by the validation script, and indeed, it is odd to have a noun as copula. But how could it be annotated in a better way? I could not find similar cases in the Irish treebank (Irish has verbnouns as Welsh). Breton is typologically different in having real infinitives.

[jnivre]

I am not sure whether the validation script requires cop to be AUX, which would be too strict, because copulas are not (auxiliary) verbs in all languages, or whether it just objects to seeing NOUN in this position. It seems that this is a case where we should allow language-specific exceptions (because the test is useful for finding annotation errors in other languages). @dan-zeman Do you have an opinon on this?

[ftyers]

In the Irish treebank, is is AUX, while bí is VERB. There appear to be no examples of bí used in its citation form.

@tlynn747 any thoughts?

[jheinecke]

Back to Welsh: if copula is in its verbnoun form, it can have possessives which would be the subject if bod was inflected. Since aux should not have children, the following is flagged by the validation script, eventhough I think it is the best way to annotate:

mi  welith  hi  fy mod   (i)  yn   anhapus        ("she will see, that I am unhappy")
AFF see-3SG she my being (my) PRED unhappy

If we attach fy and i to the predicate (adjectives) we lose the relation between this circumfix possessive with the verbnoun, fy also triggers the mutation on bod which becomes mod.
Any opinions?

1	Mi	mi	PART	aff	_	2	advmod	_	_
2	welith	gweld	VERB	verb	Mutation=SM|Number=Sing|Person=3|Tense=Future	0	root	_	_
3	hi	hi	PRON	indep	Gender=Fem|Number=Sing|Person=3|PronType=Prs	2	nsubj	_	_
4	fy	fy	PRON	dep	Number=Sing|Person=1|PronType=Prs	5	nmod:nsubj	_	_
5	mod	bod	NOUN	verbnoun	Mutation=NM|Number=Sing|VerbForm=Vnoun	9	cop	_	_
6	i	i	PRON	indep	Number=Sing|Person=1|PronType=Prs	5	expl	_	_
7	ddim	dim	PART	neg	Mutation=SM	5	advmod	_	_
8	yn	yn	PART	pred	_	9	case:pred	_	_
9	anhapus	anhapus	ADJ	adj	_	2	advcl	_	SpaceAfter=No
10	.	.	PUNCT	punct	_	2	punct	_	SpacesAfter=\n

[jnivre]

On the other hand, if you don't attach "fy" to the predicate, you lose the subject relation, which is worse given UD's emphasis on core arguments. I think this is parallel to languages like Turkish, where complementation also involves nominalisation, and I think "fy" should simply be annotated as "nsubj", not as "nmod".

[jheinecke]

I agree. But there is one thing which makes me hesitate (maybe this is not a strong argument). From a purely syntactic point of few, verbnouns are nouns, so fy mod in the example above is "possessive + noun", as is fy aval (my apple). The only difference is the xpos (verbnoun vs. noun) In the latter case I attached nmod:poss(aval, fy), in the former nmod:nsubj(mod, fy). If we attach fy to the predicate, it is clearly better from a semantic point of view, but it would make a different annotation for the same (possessive + noun)...

[jnivre]

The fact that you get a different attachment for the possessive is due to the special treatment of copula constructions in UD. For ordinary verbs, "fy" would remain attached to the verbnoun, but I still think it should have the label nsubj (not nmod:subj) to indicate that it is a clause-level argument. If it helps, try to think of "fy" as attaching to the combination of the verbnoun and the predicate, rather than the predicate alone.

[jheinecke]

Indeed, for ordinary verbnouns (not inflected verbs) it is attached to the verbnoun (as object, since the possessive on verbnouns corresponds to direct objects of inflected verbs). I misunderstood the guidelines in that nsubj/obj are reserved for verbal heads. If this is not the case, I will modify it.

[ftyers]

In Turkic languages, as @jnivre said, this would be similar, the verb is a nominal form (takes accusative case to indicate that it is a complement and it takes the embedded subject in the genitive/possessive).

mi  welith  hi  fy mod   (i)  yn   anhapus  
AFF see-3SG she my being (my) PRED unhappy

o     benim   mutsuz     olduğumu         görecek
she   my      unhappy   being.1SG.ACC see.FUT-3SG

We would annotate this as:

root(görecek)
ccomp(görecek, mutsuz)
cop(mutsuz, olduğumu)
nsubj(mutsuz, benim)
nsubj(görecek, o)

nsubj/obj are for predicate heads, but predicates can be either verbal or non-verbal.

A secondary question is why you have the complement as advcl instead of ccomp?

[jheinecke]

it's an error ... it should be a ccomp of course (I have to rethink my validation scripts, to find annotation errors from late-night annotations :-). Thanks for spotting it

[tlynn747]

Sorry to arrive late into this conversation! I can't really advise on how to label the Welsh data as I don't speak Welsh, and haven't fully followed the explanation given in order to get a clear insight. I can only explain what we've done for Irish! As Fran points out, the copula 'is' is tagged as AUX (and will be attached to the predicate with the cop label) # sent_id = 7 # text = Cailín is ea í (She is a girl) 1 Cailín Cailín PROPN Noun Case=NomAcc|Gender=Masc|Number=Sing 0 root _ _ 2 is is AUX Cop PronType=Rel|Tense=Pres|VerbForm=Cop 1 cop _ _ 3 ea ea PRON Pers Number=Sing|Person=3 4 nmod _ _ 4 í í PRON Pers Gender=Fem|Number=Sing|Person=3 1 nsubj _ SpaceAfter=No 5 . . PUNCT . _ 1 punct _ _ and the substantive form 'bí' - along with its inflected forms - are tagged as VERB, and are usually the root of the clause/ sentence. # sent_id = 901 # text = Bhí an lá an-te (the day was very hot) 1 Bhí bí VERB PastInd Form=Len|Mood=Ind|Tense=Past 0 root _ _ 2 an an DET Art Definite=Def|Number=Sing|PronType=Art 3 det _ _ 3 lá lá NOUN Noun Case=NomAcc|Definite=Def|Gender=Masc|Number=Sing 1 nsubj _ _ 4 an-te te ADJ Adj Degree=Pos 1 xcomp:pred _ They both translate into English as what is recognised as a copula but from a syntactic perspective in Irish, it's important that both are treated differently. The copula is more of a linking element, the substantive verb functions as a just that - a verb. Word order is extremely different for both. With respect to verbal nouns in Irish, there is *no verbal noun form of the copula *(such as the Welsh example of "my being") so we haven't encountered this issue of a NOUN being attached with the 'cop' label. Verbal nouns are POS-tagged as NOUN, along with morph features VerbForm=Vnoun. They typically function as infinitives (preceded by infinitive particle 'a') or gerund forms in progressive aspectual phrases, preceded by 'ag'. While the substantive verb "to be" has a verbal noun form, it is not used in a gerund manner, however. e.g. *ith *eat *a ithe* - to eat *ag ithe - *eating *bí - *substantive form of verb "to be" *a bheith - *to be **ag bheith* - this is not a valid construction, in Irish the sentence would be rephrased from "being" to "to be" to make it syntactically plausible. E.g. being happy is important to him -> it is important to him *to be happy* Teresa Ar Máirt 9 Aib 2019 ag 17:37, scríobh Francis Tyers < notifications@github.com>:

…

In the Irish treebank, *is* is AUX, while *bí* is VERB. There appear to be no examples of *bí* used in its citation form. @tlynn747 <https://github.com/tlynn747> any thoughts? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#620 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGyMg6-w9KsmcJ2m5uMvSwalBqVRQcPsks5vfMHOgaJpZM4ciNXE> .

-- *Slán agus Beannacht*

[jheinecke]

@dan-zeman, so would it be possible to have validation.py accept the verbnoun bod/ with upos NOUN as cop. Maybe not generally, but at least for languages where it makes sense, like Welsh, as @jnivre has proposed above?

[dan-zeman]

@jnivre : On the other hand, if you don't attach "fy" to the predicate, you lose the subject relation, which is worse given UD's emphasis on core arguments. I think this is parallel to languages like Turkish, where complementation also involves nominalisation, and I think "fy" should simply be annotated as "nsubj", not as "nmod".

So what is the criterion to rule that a dependent of a deverbal noun is nsubj and not nmod? I suppose there must be something else than just saying "this is the nominal dependent that plays the same semantic role as the subject would play if the head was still a (finite active) verb". Because not only UD annotates dependents of nouns differently from dependents of clauses; UD also tries not to rely on semantic roles, right? I agree that it would be useful for downstream applications to know that an nmod within a noun phrase corresponds to the subject in a clause. But if we do that, then we will have subjects under every eventive noun, in all UD languages.

[dan-zeman]

@jheinecke : Why don't you tag bod AUX instead of NOUN? That would fit in the system, and the validator would be happy, too.

[jheinecke]

Because syntactically bod is a noun and functions as such (.i.e. it can have possessives, genitives, ...)

[dan-zeman]

But if you want to attach it via a cop relation, you give it a syntactic analysis that is very different from other nouns and you definitely cannot attach a possessive or genitive directly to it. You cannot have your cake and eat it too :-)

[dan-zeman]

(BTW using the AUX tag will not prevent you from also using the VerbForm=Vnoun feature, so the information will not be lost.)

[jheinecke]

I agree, ...
Assuming I keep bod as noun, all sentences with bod + yn + predicative ADJ/NOUN have to be annotated differently from sentences like mae 'r afal yn goch (lit. is the apple PRED red). I will think about it. Maybe AUX for v2.4 and a changed analysis later on

[jnivre]

Agreed. Note that the AUX tag in UD does not imply that the word is a verb, only that it is used to in support of a predicate and carries features associated with verbal inflection (in this case only person, if I understand correctly).

[jnivre]

@dan-zeman The criteria for distinguishing clausal from nominal structures are, as in many other cases, language-specific. In the case of Turkish, the main argument is that the nominalisation is grammaticalised and obligatory in some constructions (notably complement clauses). In other languages, we have a cline such as in the standard English examples:

1. we saw that they destroyed the factory
2. we saw them destroy the factory
3. we saw them destroying the factory
4. we saw their destroying the factory
5. we saw their destroying of the factory
6. we saw their destruction of the factory

Everyone agrees (I think) that we have a clause in 1 and a noun phrase in 6. But what about 4, where the "subject" takes the form of a possessive, while the "object" is still treated as a direct object of the verb.

Bill Croft (drawing on typological studies) has proposed a criterion, which unfortunately only works in one direction. If the predicate is marked for case, then treat it as a noun (and its dependents as nmod). However, this does not tell us what to do when the predicate is NOT marked for case. A rule of thumb is to say that lexical derivations (like "destruction") should be treated as nouns, while forms that are clearly recognised as verbal (like the infinitive "destroy") should be treated as verbs. However, this still leaves a gray zone of gerunds, participles, etc. In addition, we need to consider the larger grammatical systems, like the obligatory nominalisation in complement clauses in Turkish. Exactly where Welsh verbnouns fit into this picture is difficult for me to say, but these are the general considerations, I think.

[jheinecke]

In Welsh these kind ob subordinates are always nominal (with at least the "direct object" in a genitive construction, there is rarely a conjugated verbform, simultaneity, anteriority or posteriority with respexct to the main phrase, is epxressed by TAM-markers. Closer to 1 would be (possibly not totally grammatical)

Gwelson nhw  yn   dinistrio          'r   ffatri
we-saw  they IMPF destroy-VNOUN (of-)the factory 

Closer to 6 is the (more natural)

Gwelson eu    bod      yn   dinistro      'r  ffatri
we-saw  their be-VNOUN IMPF destroy-VN of-the factory 

Gwelson eu    bod      wedi dinistro      'r  ffatri
we-saw  their be-VNOUN ANT  destroy-VN of-the factory (we was that they have destroyed the factory)

In any way the nominal structure of verbnouns is the main feature of Welsh. But in case of the verbnoun as copula, the tag AUX seems to be for the time being the best option.

[ftyers]

@jnivre +1 and great examples. @jheinecke I agree, for now this seems like the best option.

[Stormur]

@dan-zeman The criteria for distinguishing clausal from nominal structures are, as in many other cases, language-specific. In the case of Turkish, the main argument is that the nominalisation is grammaticalised and obligatory in some constructions (notably complement clauses). In other languages, we have a cline such as in the standard English examples:

1. we saw that they destroyed the factory

2. we saw them destroy the factory

3. we saw them destroying the factory

4. we saw their destroying the factory

5. we saw their destroying of the factory

6. we saw their destruction of the factory

Everyone agrees (I think) that we have a clause in 1 and a noun phrase in 6. But what about 4, where the "subject" takes the form of a possessive, while the "object" is still treated as a direct object of the verb.

Bill Croft (drawing on typological studies) has proposed a criterion, which unfortunately only works in one direction. If the predicate is marked for case, then treat it as a noun (and its dependents as nmod). However, this does not tell us what to do when the predicate is NOT marked for case. A rule of thumb is to say that lexical derivations (like "destruction") should be treated as nouns, while forms that are clearly recognised as verbal (like the infinitive "destroy") should be treated as verbs. However, this still leaves a gray zone of gerunds, participles, etc. In addition, we need to consider the larger grammatical systems, like the obligatory nominalisation in complement clauses in Turkish. Exactly where Welsh verbnouns fit into this picture is difficult for me to say, but these are the general considerations, I think.

Sorry for commenting only now.

Wouldn't it be possible to introduce a double POS (where order matters), in the sense of an "external" vs. "internal" part of speech? For example, the case you mention of destroying: we might assign it the double POS NOUN/VERB, meaning that:

• "externally" it depends from other nodes with the deprels used for NOUNs, and it can also take determiners or adjectival modifiers and so on;
• "internally" it can govern a verbal clause, in this case e.g. with an obj, but also obls, advmods, and so on.

So your sentence no. 4 could be analyzed as:

#we saw their destroying the factory
1	We	we	PRON	_	_	2	nsubj
2	saw	see	VERB	_	_	0	root
3	their	their DET	_	_	4	det
4	destroying	destroy	NOUN,VERB	_	_ 2	obj
5	the	the	DET	_  _	6	det
6	factory	factory	NOUN	_	_	4	obj

Don't we have a "bidirectional" behaviour that can be expressed like this in these cases?
Or maybe, if the VerbNoun feature is considered enough, we might allow the double nature of the deprel to/from the affected token?

[dan-zeman]

Agreed that some words sometimes behave like one category in relation to their parents and like another category in relation to their children, but no, it is not possible to express this within the UD framework. Here it is always assumed that each word in context belongs to exactly one of the 17 coarse-grained categories. If there is a secondary candidate category, it can be sometimes expressed using the features (like the VerbForm feature, which does not require UPOS to be VERB), but one category must be chosen as the main one.

[Stormur]

Does the framework consequently also not allow e.g. a NOUN token with the VerbForm feature to take verbal deprels?

Unfortunately, this brings forth lots of very weird analyses, like case-governing prepositions attached as marks to participial forms in an attributive function... Sorry if I insist, but is this really something that cannot be introduced in UD, a future milestone? It seems to be a very debated topic indeed.

[dan-zeman]

This remains somewhat controversial. Since UD makes such a strict distinction between annotation of clauses vs. nominals, I believe that if you want verbal relations, you should tag the word VERB. If it looks more like a nominal because of case marking, articles etc., it may be a derived NOUN but then it should abandon clausal relations and use the nominal ones. This way it feels like a coherent system. Note that in the scale described by Joakim, once you decide where is your boundary between clauses and nominals, you could express the boundary by using VERB (plus clausal deprels) on one side, and NOUN (plus nominal deprels) on the other side.

On the other hand, the current guidelines are less exact about this, leaving some room for diverging interpretations (which is unfortunate).

[sylvainkahane]

The inflection can change the distribution of a lexeme towards its governor. This has already been described by many authors (rank theory of Jespersen 1924, transfer of Tesnière 1959). This is exactly what happens in these examples. In 1, the transfer is done by the complementizer that, in 2 by the infinitive inflexion and in 3 and 4 by the -ing inflexion. In 6, it is a derivation and the lexeme must be classified as a NOUN. While in 2, 3 and 4, the lexeme is still a VERB but the inflected form is transferred and behaves as a noun towards its governor. (This includes the determiner, which is likely to be the surface-syntactic governor of the noun, as it is argued by many linguists, at least since Hudson 1984.)
The POS should concern the lexeme and not the inflected form. In @jnivre's examples, I don't think there is an ambiguity concerning the POS of the lexeme. The only problem I have is with example 5:
I don't know whether it is a transfer or a derivation, whether the lexeme is still the VERB destroy or a NOUN destroying. But this concerns the lemma, and as soon as we have decided what is the lemma, there is no possible discussion about the POS.
On the same way, the participles have the POS VERB, but behave as ADJ towards their governor. And in many languages, there is a gerundive form of the verb that behaves as an adverb towards its governor. The categorization of the inflected form is recorded in the morpho-syntactic features describing the inflexion (VerbForm=Part, Ger, etc.). Nevertheless, here the guidelines could be improved, because the list of possible values for VerbForm is quite confusing.
To come back to the initial post of @jheinecke, I am not sure to understand what must be done and what has been done in Welsh. When you say "In Welsh, the base forms of verbs are verbnouns, which functions as nouns.", you mean that the same lexeme has a base form where it behaves as a noun and some other inflected forms where it behaves as a verb. But if these forms are forms of the same lexeme, the lemma must be the same and the POS of this lemma must be the same in both forms, no? I suppose that the POS must be VERB, if there is another distributional class of lexemes which behave only as nouns.
If I understand the data in Welsh, it is a case where the transfer affects also the distribution towards the dependents (as in @jnivre's example 5). If it the case, it means that a VERB with VerbForm=Vnoun should have nmods, while in other forms, it has obj and obls.

[jheinecke]

Yes, verbnouns behave like nouns, inflected forms behave like verbs. Some dictionaries (like the Geiriadur Pryfysgol Cymru, http://www.geiriadur.ac.uk/, the lexical authority for Welsh), list verbs under the first person singular, present/future tense instead the verbnoun. But I would not go as far to say verbnouns are a derivation (as destruction wrt destroy).
I agree with @sylvainkahane's argument that the inflected forms keep the UPOS of the baseform, independently to the syntactic behaviour of the current form. This means for Welsh all verbnouns (currently POS=NOUN, XPOS=verbnoun) should be changed to POS=VERB, as the inflected forms are. This leaves only the question which dependency relations to use for arguments of verbnouns (which "look" like possessives:

"her house" nmod:poss(thŷ, ei)

form:  ei   thŷ
gloss: her  house
UPOS:  PRON NOUN
XPOS:  dep  noun

"to see her" obj(gweld, ei) or nmod(gweld, ei)`

form:  ei   gweld
gloss: her  see
UPOS:  PRON VERB
XPOS:  dep  verbnoun

but "I (will) see her" obj(gwelaf, hi)

form:  gwelaf      hi
gloss: see-FUT-1SG her
UPOS:  VERB       PRON
XPOS:  verb       indep

If we change the UPOS of verbnouns to VERB, I would go for the obj dependency relation.
How is this dealt with in other languages? Irish uses UPOS NOUN for verbnouns (which was the reason for me doing so as well in Welsh). Breton does not have verbnouns.

[sylvainkahane]

An inflection morpheme can change the realization of the arguments of the lexeme it combines with. For instance, a passive morpheme will change an initial object into a subject. I think I would have a similar treatement here and consider that the verbnoun form changes the initial object into an nmod:poss. Relation labels are names of constructions. Clearly in the case of ei gweld, the construction is the same as ei thŷ and not gwelaf hi.

[Stormur]

I have incurred again in the problematic annotation of verbal nouns, this time more with regard to the dependency relation of the verbal item, and would like to bring forth the following example from the Latin corpora:

• manus tradentis me lit. 'the hand of the me betraying' -> 'the hand of the one who will betray me' (Gospels)

tradentis is the genitive singular (m/f/n, all three are formally identical) of the present participle of trado 'to betray', so something like "betrayer/betraying". The fact that it still works as a verb is given by the presence of the core argument me 'me', the first person singular pronoun in the accusative case. If it were completely nominalized, I think it would rather look like mei tradentis, with a possessive determiner meus 'my/mine' concording in case and number.

So, given obj(tradentis,me) how does tradentis depends on manus? My objection to acl is that it would not represent the structure under exam: tenentis is neither an attribute of manus, nor an embedded clause governed by it (in particular, it isn't even a relative clause). I mean, from my understanding, acl is not the verbal equivalent of nmod, what I would choose here if we had something along the lines of

• manus traditoris mei 'the hand of my betrayer' -> nmod(manus,traditoris) det(traditoris,mei)

Indeed, in Latin the attributive construction with the participle would require concordance with the noun, here:

• manus tradens me 'the hand betraying me' -> acl(manus,tradens) obj(tradens,me)
  (tradens here is in the (m/f/n) nominative singular, as manus)

So manus tradens me and manus tradentis me are really structurally different. I was considering the analysis as the ellipsis of a nominal, namely (as a siede note, notice that this nominal is re-integrated in many translations in modern European languages) something like homo 'man', or also a determinant:

• manus [hominis] tradentis me 'the hand of the [man] betraying me'
  -> nmod(manus,hominis) acl(hominis,tradentis) obj(tradentis,me)

And everything is perfectly fine here. Under this light, then, I might conceive to use the orphan relation:

• manus tradentis me -> orphan(manus,tradentis) obj(tradentis,me)

It is to be noted that such constructions with a participle in a markedly nominal environment and taking adnominal modifiers seem to appear quite late in literature, from what I was able to get, so perhaps we can not rule out that we are in presence of a reanalysis, and then the relabeling of such forms as ADJ or NOUN is justified. In Italian, for example, the "verbal force" of present participles is all but vanished, and apart from some bookish example, this form survives only as lexicalized nouns and adjectives (corrente 'stream (f.)', from curro -> currens 'running'; impressionante 'impressive', from impressionare 'to impress', etc.).

Finally, just a comment about a similar case: I find similar problems with some usages of csubj. If I interpret:

tradens manum suam mihi tetendit

as 'the betrayer stretched his/her/its hand to me', then, if we were to keep the verbal form, this isn't really a "subject that is itself a clause".

---

To summarize, I see this issues/proposals/questions I have come up with:

1. we keep acl for such cases too, thereby extending its usage, even if maybe in an improper way;
2. we use orphan with the underlying analysis of a nominal ellipsis, bu tare left with a somewhat underspecified dependency;
3. we use this as evidence that these verbal nouns have to be reanalysed as ADJ/NOUNs, but then we might lose the rare clausal dependencies left and/or other verbal behaviours and/or multiply the lexicon with no real reason (i.e., one would then identify a tradens as a form in trado's paradigm, and another tradens as independent NOUN... the boundary might be very murky);
4. we introduce the equivalent of nmod for clauses, ncl (nominal clause)?, and save the day!
   • and this could be very well joined by the cobl relation proposed in this paper

To me, it seems that the relation roster of UD really needs some readjustments to treat verbal nouns in all their occurrences.

What are your thoughts about it?

[amir-zeldes]

I think it sounds like a nominalized participle, which acts like a noun meaning "betrayer", is in genitive (since it's possessed), and take an accuative obj dependent despite being a noun. So basically I'd vote for 3. Similar constructions actually occur in many languages, including Old Church Slavonic and Hebrew. I have argued here that it should act like a noun in its own deprel and still govern obj. That issue seems to still be open so there was no conclusive decision on these rara...

[gcelano]

manus tradentis me is equivalent to manus eius qui me tradit, and therefore the participle stands for a relative clause + its antecedent: if there were no accusative, the participle could be easily analyzed as a substantivized participle (therefore nmod). In this case, the presence of the dependent accusative reveals the ambivalent nature of the participle in all its strength (it is both adjective/substantivized adjective and verb). As a consequence, unless one introduces an elliptical node, it is clear that any solution would not do justice to it, in a way or in another (the current UD annotation system is binary in this respect: you can choose either a noun label or a verb one). This phenomenon is not rare in Latin and Ancient Greek. All in all, I would vote for acl.

[Stormur]

manus tradentis me is equivalent to manus eius qui me tradit, and therefore the participle stands for a relative clause + its antecedent: if there were no accusative, the participle could be easily analyzed as a substantivized participle (therefore nmod). In this case, the presence of the dependent accusative reveals the ambivalent nature of the participle in all its strength (it is both adjective/substantivized adjective and verb). As a consequence, unless one introduces an elliptical node, it is clear that any solution would not do justice to it, in a way or in another (the current UD annotation system is binary in this respect: you can choose either a noun label or a verb one). This phenomenon is not rare in Latin and Ancient Greek. All in all, I would vote for acl.

@gcelano , I would like to ask you, since you are clearly in a better position than me in that regard, if you have noticed different trends in the treatment of such verbal nouns in similar constructions in Classical Latin with respect to later, Patristic and Medieval Latin. This might help in assessing the best annotational strategy. I wonder if such difficulties simply come from trying to analyse two different varieties based only on the earlier one, when in fact new phenomena might have arisen. For example, in the corpora I have found a sentence by Palladio (XVI c.) who clearly uses currens as a modern Italian would, i.e. as a full-fledged noun, and as maybe an Ancient Roman would never had.

---

By the way: maybe the use of acl:relcl in such "not concording" cases, if the analysis of an underlying antecedent for a relative is indeed accepted, might help differentiating these cases from "more regular" acls...

[gcelano]

What was originally a participle can at a certain point become an adjective or a noun: in this case, treat it as an adjective or a noun, respectively (so the morphological annotation would not be participle anymore). Of course, there may be borderline cases: in that case, choose an interpretation and, ceteris paribus, always stick to it.

[dan-zeman]

If you tag tradentis as a VERB, then tradentis me is a clause, and since it modifies a NOUN, it must be acl. (You may want to define and document a new subtype of acl if you want to distinguish it from the other subtypes of acl.)

If you tag tradentis as a NOUN, then you cannot use the obj relation for me because UD does not distinguish core and oblique dependents of nominals. You would have to use nmod for me. But I think that participles usually become ADJ rather than nouns (note that you can still use VerbForm=Part even with ADJ), and here the picture is less sharp; I think that we have not completely ruled out direct objects of adjectives, @jnivre?

If you interpret the example as an ellipsis (manus [hominis] tradentis me), then you should use the default UD approach to ellipsis, which is promotion of one of the orphaned dependents. Here, hominis has only one dependent, tradentis, which will be promoted, i.e., it will be attached to manus as nmod. No orphan relation will be used. The orphan relation is not an equivalent of ExD in the Prague annotation style. It is used to connect one orphan to another orphan (impossible here, as we have only one orphaned dependent), and only if the elided parent was a predicate and the orphans are its arguments or adjuncts (does not apply here because hominis is a nominal).

Out of these three options, I would probably lean towards the first or the second, but I'm not strictly against the third one either. Which one should be used is the question of the broader situation in Latin (as well as the diachronic consideration you mentioned), and I'm afraid I cannot advise on that.

[jnivre]

Thanks, Dan. This looks like an excellent survey of the options. I agree that objects of adjectives may be borderline acceptable, especially for participles in languages where these are generally tagged ADJ, but I think a VERB+obj or NOUN+nmod analysis would be preferable if possible.

[passarom]

It’s always a matter of lemmatization. If you lemmatize tradentis under trado (VERB), the only option in UD is acl. At least in the IT-TB, apart from some exceptions, participles are lemmatized under the corresponding verb, following a quite strictly (inflectional) morphological criterion. In this specific case, I do not think that there are enough good reason in favor of not lemmatizing tradentis under trado, also considering that, as correctly pointed out, here the participle behaves ‘verbally’, as demonstrated by the direct object in the accusative case (me). I understand that it is not a perfect solution (but, anyway, a good one), but it is the best (and clearest) we have so far in UD. Best, Marco

…

Il giorno 10 lug 2020, alle ore 10:15, Joakim Nivre ***@***.***> ha scritto: Thanks, Dan. This looks like an excellent survey of the options. I agree that objects of adjectives may be borderline acceptable, especially for participles in languages where these are generally tagged ADJ, but I think a VERB+obj or NOUN+nmod analysis would be preferable if possible. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.