The parent of 'orphan' should normally be 'conj' but it is 'reparandum'

[amir-zeldes]

Hi - a recent update to the validator creates the error message in the title, however in the Coptic corpus we have an exception that looks correct to me: a reparandum consisting of two dependents whose head is missing. I'll give the example in English for simplicity:

... they hav-- every community according to our...
reparandum(community,they)
orphan(they,have)

The alternative of saying that both 'they' and 'have' are reparandum is unappealing, because there is a whole interrupted phrased ("they have") which results in a single repair. The option of treating 'they' as the subject of 'have' is not available in Coptic, since the equivalent of have is a past auxiliary which never takes the subject directly (it is aux). Basically there is a missing verb that would have dominated "they have", so in its absence we've promoted the subject, and treated the auxiliary as an orphan.

Any suggestion is appreciated, but if there isn't a good reason to reject orphan I would suggest allowing reparandum as a parent of orphan.

[dan-zeman]

I admit that when adding the validation rule, I assumed that it would spark discussion and possibly it would have to be loosened. Though the example you give is quite beyond my imagination :-) I have moved the issue to the docs repository because it is about precise interpretation of the guidelines (as most validation issues). The validator just tries to make sure that people can make assumptions about the data, if the assumptions follow from the documentation. The immediate impulse to introduce this rule was when I saw an orphan attached to another orphan (people probably forgot that the promoted orphan should not be attached via this relation).

In my understanding, the orphan relation was introduced solely for gapping and stripping. That is, the missing node is a predicate and the orphaned dependents are its arguments or adjuncts. It typically occurs in coordination, so the promoted argument is attached to the first predicate as conj; but the validator would also accept parataxis (which some people use instead of coordination), root (because the source can be in the previous sentence) and even advcl (because subordination can generate similar situations). It seems quite plausible that a similar pattern could occur also with reparandum, so I think I should add it. But I am not convinced that your particular example should involve the orphan relation. The guidelines assume that if the main verb is elided and an auxiliary remains, then the auxiliary is promoted and the other dependents are attached to it as if it was the main verb; which would result in the relation nsubj(have, they), which you reject.

[dan-zeman]

I whitelisted reparandum in UniversalDependencies/tools@e817893. Errors of this type are currently not reported in Coptic. I will leave this issue open for a while so that others can contribute to a better understanding of the orphan relation.

[amir-zeldes]

Thanks! The idea of promoting the auxiliary makes a lot of sense in a language like English, where the auxiliary is itself a verb. In the Coptic example, the element in question is really just a functional auxiliary, with no chance of being used as a verb, so it seems a little stranger to promote it, rather than a core argument. In terms of parallels elsewhere in the corpus, orphan is often a dependent of the subject, which get first choice as the argument to promote - for that reason I would be inclined to keep the subject as the head and say that it governs all other dependents of the missing verb - in this case also the tense marker.

[jnivre]

The point is not that the auxiliary can be used as a main verb but that it belongs to the same nucleus (in Tesnière's sense) as the main verb. As long as there is something left of the verb group, we prefer to let this represent the verb group so that other dependents can retain their true dependency relations (rather than "orphan"). It is for the same reason that we, for example, promote determiners to head elliptic noun phrases even though they can never head an ordinary noun phrase.

[amir-zeldes]

Both subject and auxiliary are dependents of the (missing) main verb - in my opinion the question is only which would we rather promote. Either way some information will be lost:

1. If we promote the auxiliary, we gain the information that the subject was probably a subject, but we lose the information that ellipsis has taken place (as marked by orphan)
2. If we promote the subject, we lose its function (since it must now bear the reparandum relation), but we retain the information about ellipsis, since its dependent will be orphan, the same relation used in other cases where a subject stands in for a missing verb

One problem with 1. for Coptic is that auxiliaries aren't present in all tenses, so we would end up with situations in which we have ellipsis and a. we do have orphan, but aux is the head; b. we do have orphan, and nsubj is the head (since there is no aux) and c. where there is no orphan at all. Promoting the subject uniformly seems like a better choice for the data we have.

I agree that in languages where the auxiliary is a finite verb (like English) it is more intuitive to promote the auxiliary, but in this case it seems like more information would be discarded, and a very odd government pattern would result (AUX ->nsubj PRON, which is impossible in Coptic, and with no trace of an orphan)

[dan-zeman]

@amir-zeldes : Just a note – orphan is not a means to mark ellipsis. Most instances of ellipsis are simply lost in UD (that is, they are hidden due to promotion). The purpose of orphan is to avoid certain relations that only occur in some instances of ellipsis and that would be "too strange". So if nsubj(AUX, PRON) is "too strange" in Coptic, then this is possibly the argument in favor of orphan. (But then we would have to document it. The aux relation is not listed in our obliqueness hierarchy, for example.)

[sylvainkahane]

To come back to the main topic of this issue, non-constituent conjunct are quite common with reformulations. Examples:

it is a good a very good question
you said something about the about my question
I think that you that we must go

If we use reparandum for such cases, the reparandum phrase (in italic in the examples) certainly needs an orphan relation.

[dan-zeman]

@sylvainkahane I think that the three examples you gave would be solved without orphan according to the UD guidelines. Simple promotion of one of the orphaned dependents.

det(good-4, a-3); det(question, a-5); advmod(good-7, very); amod(question, good-7); reparandum(good-7, good-4)

case(the, about-4); case(question, about-6); det(question, my); reparandum(my, the)

mark(you-4, that-3); mark(go, that-5); nsubj(go, we); reparandum(we, you)

[lauma]

#635 (comment) :

In my understanding, the orphan relation was introduced solely for gapping and stripping. That is, the missing node is a predicate and the orphaned dependents are its arguments or adjuncts. It typically occurs in coordination, so the promoted argument is attached to the first predicate as conj; but the validator would also accept parataxis (which some people use instead of coordination), root (because the source can be in the previous sentence) and even advcl (because subordination can generate similar situations). It seems quite plausible that a similar pattern could occur also with reparandum, so I think I should add it. But I am not convinced that your particular example should involve the orphan relation.

If advcl is okay, is acl any different? If I understand correctly, it can be a subordinate clause with its own potentially gapped predicate the same as advcl?

E.g., in Latvian saying viņš ēd tos ābolus, ko pirms tam [ēda] tārpi ('he eats the same apples, which where [eaten] by worms before that') is rather plausible.