Relative clause markers in UDv2

[nschneid]

There has been an ongoing discussion about the boundary between PRON and SCONJ for grammatical words that signal a relative clause, e.g. English "that".

Many syntacticians would argue that "that" is a pure relativizer, which would correspond to SCONJ. A discussion in #223 (2015) acknowledged this but decided to adopt the relative pronoun analysis, which allows "that" to bear an informative deprel. @jnivre wrote:

I agree with @manning. For languages like Swedish and English, treating that/att as a relative pronoun is an extremely practical solution as long as we haven't worked out detailed guidelines for the enhanced representation, and it is not obviously wrong from a linguistic point of view either despite alternative proposals. At the same time, we still want to allow other languages to treat the words introducing relative clauses as pure complementisers, when there is no evidence of relative pronouns in the language at all. As far as I understand, this is what is currently being done for Persian, for example. Finally, this will definitely be one of the issues that we need to look into when developing v2 of the guidelines, which will hopefully also include guidelines for the enhanced version.

The relative pronoun treatment is consistent with the Penn Treebank's approach—i.e. relative "that" and "which" are treated the same. (Though there are other differences: as a constituency-based framework, they use an empty element 0 for zero relatives, and make the gapped position explicit with a trace.)

The discussion resurfaced in 2021 in #475, where the guidelines were clarified.

https://universaldependencies.org/u/pos/SCONJ.html currently says:

In the case of relativizers, the tag SCONJ (and the relation mark) should only be used for words that mark a relative clause without fulfilling a syntactic role in the clause. Relative and resumptive pronouns (like [en] that and which) should be tagged PRON.

Question for discussion: Are there heuristics that can help elucidate, crosslinguistically, whether a word is a relative pronoun or a pure relativizer—i.e., whether it fulfills a syntactic role in the clause?

[amir-zeldes]

Thanks for opening this issue - let me reiterate what I feel are the main criteria for distinguishing a relative PRON from a subordinating conjunction, which are basically agreement and case marking. I feel either of them is a sufficient criterion for PRON.

The main characteristic of PRONs is that they stand in for a referring expression or participant (notwithstanding pro-verbs etc.). The key feature of a relative pronoun (and not a relative conjunction or SCONJ) is therefore the representation of the matrix nominal in the subordinate clause. If we can tell that the element under discussion is standing in for the relativized participant, then it should not be labeled SCONJ/mark and should get the dependency function that the participant is filling in the subordinate clause.

Agreement can take various forms, but is relatively easy to identify: it can be grammatical gender or number (for example classical Arabic allāði is masculine singular), or it can be animacy (for example English "who"), etc. If the relativizer carries categories corresponding to the matrix noun, then it is representing it directly. While English "who" satisfies this criterion, English "which", which can be used for both animate and inanimate antecedents, does not satisfy it (but is still a PRON due to the next sufficient condition). Classical IE languages like Ancient Greek satisfy this criterion as well, but also have case marking (e.g. "hon" indicates an accusative role in the nested clause for an antecedent which is masculine and singular).

Case marking can be straightforward, as in Ancient Greek or also in English "who" vs. "whom" (when distinguished) - the latter represents the matrix element in clauses where it is a non-subject. Again, "which" does not do this, however "which" can receive case marking in the UD sense, since it is compatible with prepositions, both in the so-called pied-piping construction ("the movie about which we spoke") and the preposition stranding construction ("the movie which we spoke about"). In both cases, we can assume "which" is a PRON, because otherwise the preposition "about" has no nominal to mark for case (or no nominal to govern in traditional phrase syntax formalisms or dependency formalisms like SUD). I'm deliberately minimizing discussion of relative "that" since I don't want the oddities of English to distract us too much, but I believe the presence of "that" in the stranding construction is enough to include it in the PRON family as well.

Typologically, languages with relative conjunctions don't exhibit any kind of case marking on the relative element. In cases where a preposition is required by the nested clause predicate, the relativizer cannot be combined with the preposition, and instead we see either zero (and no preposition, the case marking is inferred from context, e.g. in Japanese) or a separate resumptive pronoun is required, showing that the relativizer is just a conjunction. The latter is true for languages like Hebrew, Coptic and Mandarin. Here are renditions of "the man I went with" in those languages, using "what" to render an SCONJ relativizer:

1. ha ish she halaxti it o "(lit.) the man REL went.1SG with him" (i.e. "the man what I walked with him")
2. p rōme ent a i mo'she nnma f "(lit.) the man REL PAST I went with him" (i.e. "the man what I have walked with him")
3. wo gen ta yiqi zou de na ge ren "(lit.) I with him together walk REL that CLF person" (i.e. "the man what I walked together with him")

I will also add that in 3., this is probably not the most natural choice, and like Japanese, Mandarin is quite happy to use a no-pronoun relative and let an adverb like "together" express the comitative semantics ("the man what went together"), but if a preposition appears, then the pronoun has to appear too. This shows that the relativizer in all of these languages (she, ent, de) is not capable of governing the preposition/being case marked by it. I think if a language has a relativizer like that, which shows no agreement with the matrix noun and is incompatible with case marking (incl. adpositions), then it should be tagged SCONJ and attached as mark. If case marking or agreement are in evidence, it should be PRON (with PronType=Rel).

[mr-martian]

According to these criteria, I'm curious how you would handle Ancient Hebrew 'asher in obliques.

In most instances, 'asher matches your she examples, but you can occasionally get prepositions:

paqad 'et sarah ka 'asher 'amar "visited ACC Sarah as REL said" (he visited Sarah as he had said [he would], Gen 21:1)

chasach mimeni me'umah ba 'asher 'at 'ishto "you-withheld from-me something in REL you his-wife" (you have withheld something from me, in that you are his wife, Gen 39:9)

[nschneid]

While English "who" satisfies this criterion, English "which", which can be used for both animate and inanimate antecedents

Minor point but I just want to point out that "which" cannot freely be used for animate antecedents:

• Amir, who/*which works at Georgetown, made an argument about agreement and case.

So I think it exhibits agreement in at least some contexts.

[mr-martian]

I suppose I should mention what I do with the Ancient Hebrew constructions.

paqad 'et sarah ka 'asher 'amar "visited ACC Sarah as REL said" (he visited Sarah as he had said [he would], Gen 21:1)

I treat these as relative clauses with elided head nouns and promote the preposition, so this is
obl(paqad/visited, ka/as)
acl:relcl(ka/as, 'amar/said)
mark('amar/said, 'asher/REL)

[amir-zeldes]

In most instances, 'asher matches your she examples, but you can occasionally get prepositions:

No, this would not satisfy my criterion, since the preposition does not govern/case mark the participant being relativized; it is governing/marking the clause itself - "as that" and "in that" are modifiers of the clause itself, not of a participant within the clause. Relative and similar clauses can be embedded into other structures, but I would not say that is part of the diagnostics for whether the clause structure itself has a relative conjunction.

Minor point but I just want to point out that "which" cannot freely be used for animate antecedents

In general this is true, but if we look at all forms of English, it's only a strong tendency, and not a form of agreement proper in all varieties IMO. I believe I gave three counter-examples in some earlier incarnation of this issue, though some people contested whether the antecedent was truly animate or "construed as not animate" or something like that. One variety in which this definitely occurs is "Biblical English", for example from the KJV:

• The people which sat in darkness saw great light (Matthew 4:16)
• And those which remain shall hear, and fear (Deut. 19:20)

You can also find it colloquially on the Internet - here are some GUM examples:

• "Several rental places also give you the option of trained guide, which can both provide information about the sights you visit, and make sure you are safe" (I think the antecedent is the guide, not the option)
• "those who want to be scarers (which incidentally, only seem to come from two schools)" (referring to the anthropomorphic monsters in the movie Monsters Inc., but I think they can be considered animate/human-like)

TBC, I don't think it's common at all, and these could be judged as fluency errors, but GUM is also not a huge corpus, so I think it probably does occur, at least marginally. But if we want to say "which" does have inanimacy agreement that's fine too - it already qualifies as a PRON based on the case marking criterion.

[sylvainkahane]

Before discussing the POS of relative markers, I would like to do some remarks about the annotation of relative clauses.

First, to be a relative is a property of the clause, that is of the unit itself, not of the relation between this unit and its governor. I don't think it a good idea to indicate that the unit is relative clause with acl:relcl. It would be better to indicate this by a feature on the head of the clause. For instance, ClauseType=Rel.

Second, it is problematic to indicate what is the extracted position by a relation to the relative pronoun. Because some relative clauses do not have a relative pronoun and have only a gap or an ordinary pronoun. We create a catastrophe between relative clauses with a relative pronoun filling the extracted position and relative clauses with a gap or an ordinary pronoun A more homogeneous solution would be to introduce a feature indicating the extracted syntactic position, for instance by a feature Filler=Obj if the extracted position is obj. This feature could be positioned on the governor of the extracted position. It would be quite similar to an enhanced dependency, but simpler and lighter.

Example the person I hope to meet
Additional features: hope: ClauseType=Rel; meet: Filler=Obj
Possible change: acl(person,hope), rather than acl:relcl

The situation is even more problematic for free relatives, where the relative pronoun is analyzed as the head and therefore the extracted syntactic position is never identified. Moreover, the acl:relcl is between the relative pronoun and the rest of the clause, while the relative pronoun is part of the free relative. See examples in English. Here it would be much better to have a feature ClauseType=FreeRel on the head of the relative clause (rather than acl:relcl) and a feature Filler indicating the extracted position.

Example: See what this says.
Additional features: what: ClauseType=FreeRel; says: Filler=Obj.

If I says this before answering the question about the PRON vs SCONJ analysis of the relative markers, it is because I think that favoring the analysis of relative markers as PRONs was also motivated by a desire to indicate the extracted position.

[nschneid]

May I ask why you suggest a syntactic Filler feature rather than an enhanced dependency? Thus far the features in FEATS have been limited to morphology. (We stretch that a bit for ExtPos but that has to do with the relation of the word's category and the phrase's category in special situations.) If we did not have enhanced dependencies then I agree there would need to be another place for this information.

On Tue, Dec 30, 2025, 2:15 PM Sylvain Kahane ***@***.***> wrote: *sylvainkahane* left a comment (UniversalDependencies/docs#1192) <#1192 (comment)> Before discussing the POS of relative markers, I would like to do some remarks about the annotation of relative clauses. First, to be a relative is a property of the clause, that is of the unit itself, not of the relation between this unit and its governor. I don't think it a good idea to indicate that the unit is relative clause with acl:relcl. It would be better to indicate this by a feature on the head of the clause. For instance, ClauseType=Rel. Second, it is problematic to indicate what is the extracted position by a relation to the relative pronoun. Because some relative clauses do not have a relative pronoun and have only a gap or an ordinary pronoun. We create a catastrophe between relative clauses with a relative pronoun filling the extracted position and relative clauses with a gap or an ordinary pronoun A more homogeneous solution would be to introduce a feature indicating the extracted syntactic position, for instance by a feature Filler=Obj if the extracted position is obj. This feature could be positioned on the governor of the extracted position. It would be quite similar to an enhanced dependency, but simpler and lighter. Example *the person I hope to meet* Additional features: *hope*: ClauseType=Rel; *meet*: Filler=Obj Possible change: acl(person,hope), rather than acl:relcl The situation is even more problematic for free relatives, where the relative pronoun is analyzed as the head and therefore the extracted syntactic position is never identified. Moreover, the acl:relcl is between the relative pronoun and the rest of the clause, while the relative pronoun is part of the free relative. See examples in English <https://universal.grew.fr/?custom=69541fdc3ad67>. Here it would be much better to have a feature ClauseType=FreeRel on the head of the relative clause (rather than acl:relcl) and a feature Filler indicating the extracted position. Example: *See what this says.* Additional features: *what*: ClauseType=FreeRel; *says*: Filler=Obj. If I says this before answering the question about the PRON vs SCONJ analysis of the relative markers, it is because I think that favoring the analysis of relative markers as PRONs was also motivated by a desire to indicate the extracted position. — Reply to this email directly, view it on GitHub <#1192 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAHQRL3OK4KDQNKOKITN7C34ELFOVAVCNFSM6AAAAACQH4HN52VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTOMBQGIZDSMZRGI> . You are receiving this because you authored the thread.Message ID: ***@***.***>

Nathan Schneider Associate Professor of Linguistics and Computer Science, Georgetown University http://nathan.cl

[sylvainkahane]

May I ask why you suggest a syntactic Filler feature rather than an
enhanced dependency?

Because enhanced dependencies belong to another formalism. Let's take the first example of a free relative in English-EWT: To whom it may concern. The Filler solution means to add the feature Filler=Obl on concern. Everybody annotating in UD can do that. The enhanced dependency solution means adding 5:obl:to in the 9th column. It is a new formalism to learn and I am not sure that all annotation tools support this.

[sylvainkahane]

I agree with @amir-zeldes that "the main criteria for distinguishing a relative PRON from a subordinating conjunction, which are basically agreement and case marking." Where I disagree is considering preposition stranding as case marking. Case marking is a dependent-marking construction. But there are also head-marking constructions (see (see for instance the remarkable article of Johanna Nichols 1986 I have already mentioned). The most clear head-marking is when the marker is a suffix on the governor, as in Wolof:

xaj bi ‘the dog’, lit. dog the
xaju Peer bi ‘Peter's dog', lit. dog.of Peter the

When we have an adposition, we have a marker of dependency that floats between the governor and the dependent. We tends to consider that this marker is closer to the dependent than to the governor, because this marker generally move with its dependent. The UD analysis goes far in this direction by considering that the adposition in only linked to the dependent. But this analysis is far to be shared by all linguists and it becomes very problematic when the language shows head-marking properties, as it is the case in English with prepositions. Let me recall the numerous clues we have that in English the preposition as a strong connection with the verb (maybe stronger than with the noun depending on the verb):

• phrasal verbs (such as go on), where particles are often prepositions
• preposition stranding in relative clauses,
• in a coordination of nouns introduced by a preposition, the preposition is repeated in English in only 10% of cases. In French, it is in 50%
• passivization: Mary talked to Peter -> Peter was talked to
• gapping coordination: Peter will go to London and Mary, Paris
• tough-movement: too expensive for me to play with

None of these constructions would be possible in French, where the preposition must always stay with nouns.
For me the fact that the preposition can never never never move with the relativizer that is a clear evidence that this word does not attract the preposition. It is a clear case where the preposition is a head-marker and not a case-marker.
According to our criteria, there is no doubt that that must be analyzed as an SCONJ.

[nschneid]

The enhanced dependency solution means adding 5:obl:to in the 9th column. It is a new formalism to learn and I am not sure that all annotation tools support this.

My understanding is that the enhanced graph is UDv2's solution to the limitations of trees-over-tokens for representing syntax—including coordination, nonlocal dependencies, and predicate ellipsis. If the community finds enhanced dependencies too complex to understand and use, or inconvenient to annotate, then perhaps more concise alternatives can be considered for UDv3. But for me that's beyond the scope of this issue.

• preposition stranding in relative clauses,

• passivization: Mary talked to Peter -> Peter was talked to

• tough-movement: too expensive for me to play with

Preposition stranding is also possible in WH-questions: What are you sitting on?

[sylvainkahane]

Preposition stranding is also possible in WH-questions: What are you sitting on?

First, English wh-relativizers allow preposition stranding, but they also allow the preposition to move with them (the so-called pied-piping). So we can say that wh-relativizers accept case-marking, with our extended definition of case-marking, where we consider that the preposition is a case marker when it sticks to the noun.

When the preposition sticks to the governor and not the dependent of the relation it marks, it would make sense to attach the preposition to the governor. This case is not covered by the current UD annotation scheme and a new relation is certainly needed. I don't know how this relation should be named. But in all the constructions where we have clear evidence that the marker of the relation is closer to the governor than to the dependent, we could attach it to the governor with this relation. In other words, in the three relative constructions of English accepting the preposition stranding we could have the same analysis of the preposition. The only difference would be that which would be obl, that would be mark and in case of zero marking there will be nothing (except a feature Filler=Obl or an enhanced dependency).

The same relation between a verb and a preposition would be used in the other constructions I mentioned: gapping coordination, passivization, tough-movement. Phrasal verbs are different because the link between the verb and the preposition is lexicalized and the preposition no longer marks a relation.

[nschneid]

But in all the constructions where we have clear evidence that the marker of the relation is closer to the governor than to the dependent, we could attach it to the governor with this relation.

Attaching "on" to "sitting" in "What are you sitting on?" but to "chair" in "You are sitting on the chair" would prioritize word order to a greater extent than UD generally does. UD abstracts away from word order differences in WH-questions and subject-auxiliary inversion; I don't see why prepositions should be any different.

(Moreover, among syntacticians generally, I think the stranding-with-a-gap account is the standard one; I haven't seen a treatment of prepositions as words that generally alternate between head-marking and dependent-marking in English, as governor+preposition would fail standard constituency tests. For idiomatic ones ['phrasal verbs'] there is some disagreement.)

As a reminder, preposition stranding need not have a verb as governor:

• Which car are you the owner of?
• Which paper are you an author on?

It is easy to tell annotators to consider the declarative paraphrase and supply the same UD structure.

How does SUD handle preposition stranding? Because SUD is more surface-oriented than UD I would not be surprised if it differs.

[sylvainkahane]

In almost all theoretical frameworks (X-bar syntax, HPSH, LFG, traditional DG, SUD), the preposition is analyzed as the head of the PP. In DG or SUD, it means that the preposition is between the governor and the dependent of the relation it marks. For instance:

• I talked to Ann: talked -> to -> Ann
• the owner of this car: owner -> of -> car

In consequence, in such an analysis, you don't decide (and you don't indicate) whether the preposition is closer to the governor than the dependent.

It is really a question specific to UD that analyzes the preposition as attached only to the dependent. In most constructions, it is not a problem, but when the dependent is extracted and the preposition remains with the governor it becomes problematic. For these cases (and only for these), we need a way to attach the preposition to the governor.