home edit page issue tracker
This page still pertains to UD version 1.

Core Dependents in UD v2

The distinction between core dependents and the rest is fundamental to the whole taxonomy. Having specific and cross-linguistically consistent guidelines for core dependents is therefore crucial for putting the whole enterprise on a solid footing. This involves clarifying the treatment of (among other things) double objects, reflexives, expletives, copula constructions and valency-changing operations. Relevant reports from the Uppsala meeting includes: copula, clitics.

What are core arguments?

Dan’s proposal: The UD guidelines depend heavily on the notion of core arguments, yet they lack a good definition of core arguments. (They just say that the distinction is “fairly clear”, which I disagree with.) What are the tests to distinguish a core argument from a non-core one? We say that we avoid distinguishing arguments from adjuncts because the borderline is fuzzy and semantic role labeling is difficult. Distinguishing core arguments from other arguments and adjuncts is supposed to be easier, but without the definition it isn’t.

In English UD, arguments that use prepositions are considered oblique, i.e. non-core. As a result, John is indirect object (iobj) in I gave John the book but only a modifier (nmod) in I gave the book to John, although the two sentences do not differ semantically and the syntactic difference is rather minimal (swapped word order and one preposition). The mere presence of a preposition cannot destroy the “coreness” of the argument, definitely not cross-linguistically. (After all we like to claim that adpositions and case affixes are the same thing, don’t we?) So in other UD treebanks we may encounter objects with prepositions.

Dan’s note: Natalia told me that the idea was that a core argument must use a coding strategy that is never used for adjuncts in the given language. I find it too restrictive. It may be acceptable in English (although I’m not particularly happy with the John inconsistency mentioned above). But in other languages it would wipe out core arguments completely. In Czech, even the accusative case (which is arguably the most widely used coding of a direct object) would be excluded because it is also used for temporal adjuncts. Not to mention that the Czech grammar denotes noun phrases in various other morphological cases (with or without prepositions) as objects.

If there is no good, exact and cross-linguistically applicable rule (and I would be surprised if it existed), maybe we can increase the degree of freedom here and come up with a set of less sharply described principles and recommendations? For example, if the most natural way of querying the dependent is using interrogative adverbs such as where, when, how, why, the dependent is adjunct and thus nmod or advmod (or advcl). If the query uses interrogative pronoun (who, whom, what, on what) AND an adverb is not suitable, we probably have a (core?) argument, i.e. nsubj, dobj, iobj. Hence, on Saturday would be nmod in I’m leaving on Saturday (when?) but dobj in I rely on Saturday (what on?) It is quite likely that there are tough border cases in all languages. But I like this approach better than the vacuum we have now, and definitely better than banning prepositions.

So why do I insist that prepositions are crucial? Because for a language like Swedish it gives an almost razor sharp distinction. By contrast, if we try to draw the distinction between complements and adjuncts (where prepositional phrases can be either), it is hopelessly muddled. For example, take a sentence like “she slipped on the sidewalk”. Can you say “what did she slip on?”, or do you have to say “where did she slip?” I suppose the latter is more natural in most cases, but the former definitely seems possible. So is “on the sidewalk” dobj or nmod? I would say it is clearly nmod and this falls out naturally with the ban on prepositions. So before giving up this idea, I would like to know whether there is any way of making this work for other languages too, not necessarily by banning prepositions in all languages but by looking for corresponding clues in other languages. Dan on the sidewalk example: That’s why I would only try to ask what on if I cannot ask where. Here I apparently can ask where, so it is an nmod. But I don’t say my approach is razor sharp and I guess you could eventually bring up an example that I could not solve satisfactorily. On the other hand, my approach more or less follows the way syntax is taught in Czech elementary schools, and how it distinguishes “objects” from “adverbials” (where a noun phrase can be an “adverbial” too; you could as well say arguments and adjuncts). We have this distinction in the Czech data, so I have to choose one of possible ways: 1. abandon it (but I don’t like losing information); 2. keep it, but separate from the basic UD representation (i.e. language-specific nmod:arg, enhanced or MISC); 3. adjust the Czech-specific definition of core arguments so that all our “objects” fit in the core (the current solution; but at the expense of cross-linguistic compatibility). There are phrases that are clearly adjuncts (like most temporal and locational modifiers). There are others that are less clear, which is why UD (but not Czech UD) currently avoids the argument-adjunct distinction. And there are yet others that are clearly arguments – not necessarily obligatory, whatever that means, but nevertheless their encoding and meaning is defined by the verb. Even the patient role may be encoded by non-accusative case, with or without preposition.

Joakim’s thoughts on an alternative proposal: The UD taxonomy, as I understand it, rests on the distinction between core arguments and obliques, which is commonly assumed in linguistic typology to be more surface oriented and therefore easier to maintain than the distinction between arguments and adjuncts. It starts from the assumption that all languages have some prototypical way of encoding the arguments of intransitive and transitive verbs, often referred to as S (for the subject of an intransitive verb), A (for the subject/agent of a transitive verb) and O or P (for the object/patient of a transitive verb). Exactly what this prototypical encoding is has to be established specifically for each language but it often involves some combination of case-marking (nominative-accusative or ergative-absolutive) and/or indexing on the verb (agreement) and/or linear position in the clause (typically relative to the verb). To this may be added the possibility to undergo certain grammatical transformations, such as relativization and passivization. The notion of core argument is then reserved for dependents of the verb that exhibit all or most of this prototypical encoding.

All other dependents of the verb are said to be oblique, which again may entail different things for different languages. In English, it means having a prepositional marker and/or occurring in a different position relative to the verb than core arguments. For example, “on her” in “I called on her” is oblique because it has a preposition, but “this morning” in “this morning I called on her” is oblique because it is not in subject position nor in object position and because it cannot be passivized. For case languages, obliques may either be accompanied by adpositions or occur in cases that are not prototypical for core arguments (often referred to as oblique cases). Exactly which cases are regarded as oblique can again vary between languages, and typical borderline cases are dative, partitive and (less common) genitive. Crucially, however, obliques can be arguments as well as adjuncts, as illustrated by the earlier English examples.

One issue that needs to be discussed is under what circumstances subjects can occur with oblique encoding. In Turkish subordinate clauses, for example, the verb is nominalized and the subject occurs in genitive instead of nominative. This is a completely regular grammatical process of complementation, and it occurs regardless of which verb is involved, so it seems well motivated to say that the genitive nominal is still the subject. By contrast, many languages express possession with a form of the verb “to be”, the thing owned in nominative and the possessor in an oblique case like dative. That is, instead of saying “X has Y”, they say “Y is to X”. In this case, there is no reason to say that the dative nominal is the subject, just because the corresponding nominal would be the subject in other languages. After all, it is precisely this kind of systematic difference between languages that we want to be able to study.

Finally, if we are going to (continue to) use this as the basis for the UD taxonomy, then we should probably require that the language-specific documentation describes how the core-oblique distinction has been applied for a given language and preferably also lists the prototypical encoding properties of core arguments. For example, in a language like Spanish, prepositions are used in the prototypical encoding of direct objects but only for animate objects with specific reference (“buscan una secretaria” vs. “buscan a una de las secretarias”).

Dan’s comment: Some people would say that indirect objects are not core arguments (John Sylak-Glassman, draft 2 of UniMorph). That seems a bit more consistent to me. We would distinguish just the S, A, P roles and everything else. We would have [nc]subj (=S+A) and dobj (=P) but not iobj – that would be just nmod, or alternatively, in languages where it makes sense, nmod:iobj.

Nathan: For English, the only clear-cut syntactic distinctions (it seems to me) are whether the argument is marked with a preposition or not, and if not, whether it occurs in subject/object position or is adverbial. So I’m fine with the current nominal argument distinctions of nsubj, dobj, iobj (core); nmod, nmod:tmod (oblique); and expl (not a true argument). It doesn’t bother me that an argument that is core in one language can be oblique in a paraphrase or translation (this seems inevitable). In languages with case, are there any cases that are considered borderline between core and oblique?

Dan to Nathan: Can you first define core and oblique for these languages? :) (This is the egg-and-chicken problem. My main concern was that we lack a portable definition of this borderline. Once we have it, I may possibly criticize it, pointing out what it does in individual languages. But without it, further discussion is pointless.)

Chris thoughts on “core vs. non-core”: My thoughts largely correspond to “Joakim’s thoughts on an alternative proposal”. There is considerable linguistic literature on the core/non-core distinction, which should at least be consulted. I will admit that as usual in linguistics, once you start looking at more and more languages, things become less than crystal clear and there are difficult cases. Nevertheless, I take it that there is a central notion that languages have some most basic and frequently used actant coding strategies (whether word order, case or agreement) which are differentiated from other actant codings that are more limited to particular purposes. In easier to understand words, in many language, that means that subject and object(s) - or the A, S, O roles - are core arguments and other things aren’t. The most frequently invoked claim is that core argument positions can normally hold a variety of semantic roles, whereas oblique coding strategies are more specialized for a particular semantic role. (This is a little bit connected to the question word test.) In many languages there are syntactic tests which pick out a set of “core” arguments, though most of these tests are particular to certain language families (e.g., verbs agree with them, these arguments can be relativized, these arguments don’t require a preposition, valence changing operations (passive, causative, etc.) target core arguments.) While various tests are useful in various language families, I think there can be no simple reduction to case or lack of a preposition. Traditional case names like nom/acc/erg/abs certainly usually indicate core roles, but beyond this being a bit circular, there are usually exceptions. Many languages will also use an accusative for some other purposes (such as a goal in Russian). Many languages have some objects or subjects that do not appear in nom/acc (such as quirky case in Icelandic). Sometimes we have to recognize a prepositionally marked argument as core (such as animate direct objects in Spanish). In sum, I would be happy with the definition: “Core arguments are the syntactic means a language uses to express the basic actants of a verb, i.e., A, S, O/P in the sense of Dixon/Comrie and any additional arguments that are basically treated similarly in marking and syntactic behavior (e.g., in “multiple object” languages).”

We should at least consult some of the relevant linguistic literature. Here’s a couple of papers surveying the core-oblique distinction:

Even though the latter is pointing out that the distinction can be hard to operationalize in various Indonesian languages - I guess these are just the kind of typologically hard cases that at some point that we need to think about! In part I think we have to be modest in our goals. We’re not likely to do better than the last century of modern linguistic thought (!). All these issues have been much debated there (such as the “measure nomininals” like “He weights 80 pounds” being bare NPs after a verb but not fully behaving like objects in many respects). I think we just have to make some pragmatic decisions and move on.

Chris on u-dep/iobj: I think there is some confusion on how we intended u-dep/iobj to be interpreted. It wasn’t meant to be thinking of, say, dative case in Latin, but rather for “other objects”. The claim/history is this: Beyond A, S, and O, some language families allow verbs that are “more than transitive”. They have verbs that allow 3 or more actants to be coded as core arguments, or in other words that they have 2 or more objects. Canonical cases are ditransitive verbs in Germanic languages and the ditransitive (and indeed tri-transitive) verbs of Bantu languages. Other languages lack such ditransitives altogether (e.g., Romance languages). It seems reasonable to regard these other objects as also core arguments. If we wish to have a unique u-dep/dobj - and we did - the question is which one to choose, say in a Germanic ditransitive of giving. LFG argued for the recipient/benefactive object on the basis of facts of word order, canonical passive, etc. But that has been questioned quite a bit since, by Hudson and others, and some (Bantu) languages seem to treat objects symmetrically. Overall, the arguments for which object of a ditransitive to regard as primary come out as quite mixed. We decided that it would be better for applications and less confusing for non-PhD linguist users if we made the argument that is the “proto-patient” (Dowty’s term) or the most theme/patient like one the primary object, using the traditional term “direct object” (dobj) and then used the traditional term “indirect object” (iobj) for other “other objects” – oobj would have been a bit confusing. There is a good discussion of object properties for English in Huddleston and Pullum (2002) The Cambridge Grammar of the English language, starting at p.248. They end up arguing for exactly the same position with their O^d = dobj and O^i = iobj, though neither me nor Marie had actually read this when we initially set things up the way we did.

Valency-changing operations

Joakim’s proposal: The current guidelines distinguish u-dep/nsubjpass, u-dep/csubjpass, and u-dep/auxpass from u-dep/nsubj, u-dep/csubj and u-dep/aux to capture the fact that the subject of a passive has a different role than the subject of the corresponding active verb. While I can clearly see the usefulness of this convention for many IE-type practical applications, it seems to go against the general spirit in UD of annotating syntactic functions rather than semantic roles. A possible counterargument would be that passive is a grammaticalized process and therefore part of syntax, but the counter-counterargument to this is that the same should then be true of causative, antipassive, etc. in many languages. First of all, I don’t think we want a proliferation of grammatical relations like “dobjcaus”, etc. Secondly, I think the lexicalist stance in UD actually speaks in favor of a lexicalist account of these operations (as in LFG, for example). My proposal is therefore that we get rid of u-dep/nsubjpass, u-dep/csubjpass, and u-dep/auxpass and instead treat all valency-changing operations as derivations (which we have agreed to do for Turkish causatives, for example). If we want to capture the grammatically regular shifts in valency, we can do so (optionally) in the enhanced dependencies. See examples for passive in English and causative in Japanese below. (I am unsure what relations should be used for the “deep” relations, so I just put “dobj:prom” [for “promoted dobj”] and “nsubj:dem” [for “demoted nsubj” for now].) An alternative approach would be to make u-dep/nsubjpass a language-specific subtype (thus, “nsubj:pass”). In this way, treebanks that don’t make this distinction (and I think we have a significant number) would be legitimate, and languages that don’t have a passive wouldn’t have to worry about it. Languages that have a grammaticalized causative could conceivable add “dobj:caus”, etc. This is in many ways a simpler solution, but I am afraid we would be overloading the language-specific relations with too many distinctions.

The cat chased the dog . nsubj(chased, cat) dobj(chased, dog)
The dog was chased by the cat . nsubj(chased, dog) nmod(chased, cat) dobj:prom(chased, dog) nsubj:dem(chased, cat)
The dog was chased from the room . nsubj(chased, dog) dobj:prom(chased, dog) nmod(chased, room)
Kanako ga Ziroo o ikaseta . \n Kanako made Ziro go . nsubj(ikaseta, Kanako) dobj(ikaseta, Ziroo) nsubj:dem(ikaseta, Ziroo)

Chris on -pass relations: I can totally see Joakim’s point, and would agree that the treatment of causatives was certainly not very developed in the original UD proposals but my general belief is that we should adopt the kind of lexicalist approach to causatives that he outlines. The ideas of dobj:prom and nsubj:dem would be possible for an enhanced representation, but seem too complex for UD basic dependencies. I’d be okay with losing the -pass relations. However, the big reason for putting them in was that they are very useful on practical grounds. If you are writing a relation extraction system and are using lemmas, then for most verb forms picking out the path “nsubj” will consistently pick out one semantic role (say, the agent of an acquiring event), but if the verb is passivized, it then picks out the acquired company instead. Our feeling was that if you did mark -pass and you had a primary/other object distinction then in practice you had enough resources that a surface syntactic representation would almost always be enough that features for relation extraction would work well, with much less cost and less representational difficulty than attempting to build such systems on semantic roles. I still sort of feel that way. So, possible futures are: (i) Use a feature to show passive and to make sure everyone uses them (features weren’t part of SD, after all, so we needed -pass to have some representation); (ii) Allow nsubj:pass etc. as “language-particular” relations for people who like them; (iii) Attempt a useful typological definition that still has -pass. If the last were to be attempted, I would do it like this: “Languages have a basic mapping of A, S, O roles onto grammatical relations, which is unmarked, common, and available to at least nearly all verbs. Many languages then have one or more syntactic processes whereby an A/S/proto-agent argument which would normally become a core role is demoted. In many but not all languages, another role then serves as syntactic subject. In such cases, this “replacement subject” is marked with a -pass to show that it is not the unmarked filler of the subject role for that verb.