Ellipsis in UD v2
The remnant relation turned out to be a non-optimal way for analyzing complex ellipitical constructions. We therefore propose the following changes for UD v2:
- The
remnantrelation should be discarded - In the case of simple head ellipsis, the dependent should be promoted
- In case of predicate ellipsis, we consider two alternative proposals:
- Use a more complex relation type that combines the relation of the elided node and its dependent (e.g.,
conj>dobj) - Use promotion but introduce a new relation
orphanfor non-standard dependency relations that arises
- Use a more complex relation type that combines the relation of the elided node and its dependent (e.g.,
- The complex cases of ellipsis should be analyzed with NULL nodes in the enhanced representation
Problems with the remnant analysis
The current analysis of elliptical constructions using the remnant relation does not work well when the second clause contains additional modifiers of the elided predicate.
They had left the company , many for good .
nsubj(left, They)
dobj(left, company)
remnant(They, many)
nmod(many, good)
In this example, for good is modifying the elided left of the second clause. However, as no similar modifier exists in the first clause, for good cannot be attached with a remnant relation and no reasonable analysis of this sentence is possible. In practice, annotators attached the extra modifier to the subject of the second clause, which incorrectly suggests that for good is modifying the subject many.
Additional issues of the remnant analysis are:
- The dependency trees contain a lot of non-projective dependencies, which complicates parsing.
- The requirement that every argument/modifier of the elided predicate has an antecedent, makes it impossible to analyze cross-sentence elliptical constructions. (This would require relations crossing sentence boundaries.)
Head ellipsis in UD v2
In the following cases of head ellipsis, we promote a modifier of the elided head.
Nominals
If the head nominal is elided, we promote dependents in the following order: amod > nummod > det > nmod > case.
Examples:
Er kauft sich ein grünes Auto und sie kauft sich ein rotes . \n He buys himself a green car and she buys herself a red .
nsubj(kauft-2, Er-1)
det(Auto-6, ein-4)
amod(Auto-6, grünes-5)
dobj(kauft-2, Auto-6)
conj(kauft-2, kauft-9)
nsubj(kauft-9, sie-8)
dobj(kauft-9, rotes-12)
det(rotes-12, ein-11)
She saw every animal at the zoo but he saw only some .
nsubj(saw-2, She-1)
det(animal-4, every-3)
dobj(saw-2, animal-4)
conj(saw-2, saw-10)
advmod(some-12, only-11)
dobj(saw-10, some-12)
She saw three monkeys and he saw two .
nsubj(saw-2, She-1)
nummod(monkeys-4, three-3)
dobj(saw-2, monkeys-4)
conj(saw-2, saw-7)
dobj(saw-7, two-8)
Clauses
If the main predicate is elided, we promote only if there is an aux or cop, or a mark in the form of an infinitival marker.
Example:
Sue likes pasta and Peter does , too .
nsubj(likes-2, Sue-1)
dobj(likes-2, pasta-3)
conj(likes-2, does-6)
nsubj(does-6, Peter-5)
advmod(does-6, too-8)
Sue is hungry and Peter is , too .
nsubj(hungry-3, Sue-1)
cop(hungry-3, is-2)
conj(hungry-3, is-6)
nsubj(is-6, Peter-5)
advmod(is-6, too-8)
They will do it if they want to .
nsubj(will-2, They-1)
aux(do-3, will-2)
dobj(it-4, do-3)
advcl(do-3, want-7)
nsubj(want-7, they-6)
xcomp(want-7, to-8)
Predicate ellipsis in Basic UD v2
In more complicated cases where a predicate is elided but no aux or cop is present, promotion would lead to very unnatural and confusing relations. For example, in the following sentence, you would be the subject of coffee, suggesting that the second clause contains a copular construction rather than an elided predicate.
I like tea and you coffee .
nsubj(like-2, I-1)
dobj(like-2, tea-3)
nsubj(coffee-6, you-5)
conj(like-2, coffee-6)
We consider two alternative proposals for dealing with such cases in UD basic dependencies, one that make use of composite relations (but do not introduce any new relations) and one that instead adds a new relation tentatively dubbed orphan to preserve intuitions about constituency.
Predicate ellipis 1: composite relations
The first alternative is to attach orphans to their grandparent with a composite relation of the form headrel>orphanrel.
Example:
I like tea and you coffee .
nsubj(like-2, I-1)
dobj(like-2, tea-3)
conj>nsubj(like-2, you-5)
conj>dobj(like-2, coffee-6)
If the grandparent is also elided, the relation is composed of all three relations and the orphan is attached to its great-grandparent.
Example:
Mary wants to buy a book and Jenny a CD .
nsubj(wants-2, Mary-1)
xcomp(wants-2, buy-4)
dobj(buy-4, book-6)
conj>nsubj(wants-2, Jenny-8)
conj>xcomp>dobj(wants-2, CD-10)
Unlike the analysis using the remnant relation, this proposal also allows us to analyze sentences in which the second clause contains additional modifiers.
They had left the company , many for good .
nsubj(left, They)
dobj(left, company)
conj>nsubj(left, many)
conj>nmod(left, good)
This approach can also be used when the antecedent of the elided node is in another sentence. The artificial ROOT node is now allowed to have more than one child, but only if all are attached via composite relations, starting with root>.
Mary wants to buy a book . ROOT And Jenny a CD .
nsubj(wants-2, Mary-1)
xcomp(wants-2, buy-4)
dobj(buy-4, book-6)
root>nsubj(ROOT, Jenny)
root>xcomp>dobj(ROOT, CD)
Predicate ellipis 2: orphan instead of remnant
The second alternative preserves the integrity of the second conjunct as a single subtree by (arbitrarily) promoting one of the orphans to the (subclause) root and attaching the others with a new dummy relation orphan (or possibly ncc for non-constituent coordination). Here are the same examples annotated according to this alternative:
I like tea and you coffee .
nsubj(like-2, I-1)
dobj(like-2, tea-3)
conj(like-2, you-5)
orphan(you-5, coffee-6)
Mary wants to buy a book and Jenny a CD .
nsubj(wants-2, Mary-1)
xcomp(wants-2, buy-4)
dobj(buy-4, book-6)
conj(wants-2, Jenny-8)
orphan(Jenny-8, CD-10)
They had left the company , many for good .
nsubj(left, They)
dobj(left, company)
conj(left, many)
orphan(many, good)
Mary wants to buy a book . ROOT And Jenny a CD .
nsubj(wants-2, Mary-1)
xcomp(wants-2, buy-4)
dobj(buy-4, book-6)
root(ROOT, Jenny)
orphan(Jenny, CD)
Predicate ellipsis in Enhanced UD v2
While we hold on to the principle that basic UD trees have to be strict surface syntax trees, we propose to relax this requirement in the enhanced representation and to allow special null nodes for sentences with elided predicates. These nodes have special word indices of the form a.b, where a is the index of the token that would precede the elided word and b is a counter. (See also the description of the proposed changes to the CoNLL-U file format.) Whenever the basic representation contains a composite relation under proposal 1 above, the enhanced representation contains additional null nodes to resolve all composite relations into simple relations.
For example, the sentences from the previous section are analyzed as following in the enhanced representation. (The special null nodes are labelled with Ea.b .)
I like tea and you E5.1 coffee .
nsubj(like-2, I-1)
dobj(like-2, tea-3)
nsubj(E5.1-6, you-5)
conj(like-2, E5.1-6)
dobj(E5.1-6, coffee-7)
Mary wants to buy a book and Jenny E8.1 E8.2 a CD .
nsubj(wants-2, Mary-1)
xcomp(wants-2, buy-4)
dobj(buy-4, book-6)
conj(wants-2, E8.1-9)
nsubj(E8.1-9, Jenny-8)
xcomp(E8.1-9, E8.2-10)
dobj(E8.2-10, CD-12)
They had left the company , many E7.1 for good .
nsubj(left, They)
dobj(left, company)
conj(left, E7.1)
nsubj(E7.1, many)
nmod(E7.1, good)
In the first example, the node E5.1 is added for the elided predicate like. In the second example, we add one node for the elided matrix verb wants (E8.1) and one node for the elided embedded verb buy (E8.2). As the elided marker to does not have any dependents, we do not add a null node for it. We also recommend adding a link between the null nodes and their corresponding surface forms in the sentence (e.g., linking E5.1 to like in the first example)?
Finally, it should be noted that, if we adopt alternative 1 for the basic dependencies, then the enhanced representation can in most cases be inferred automatically, whereas with alternative 2 additional annotation will be needed.