A prototypical relative clause (RC) modifies a nominal that is understood to fulfill some grammatical role in the RC. The head is said to be “extracted” from the RC. Occasionally the head may be another clause rather than a nominal; this special case will be discussed below.

In the basic dependency representation, the main predicate of the prototypical (that is, adnominal) RC attaches to the head as acl:relcl (if a language has relative clauses, it should use the :relcl subtype to distinguish relative clauses from other adnominal clauses). This is shown in the example on the left.

In the enhanced dependency layer, there is an additional dependency in the opposite direction to indicate the role from which the head was “extracted”. This is shown on the right.

The RC may contain a relativizer (relative pronoun that, who, relative determiner which, whose etc.; see PronType=Rel). The relativizer can be understood as an anaphor whose antecedent is the head of the relative clause. Not all languages use a relativizer. Even in English, the relativizer is optional in some contexts, as in the object relativization in the examples above.

Basic UD (left) analyzes the relativizer, if present, as filling a role in the RC. Specifically:

• Pronominal relativizers (that, which, who, what, etc.) are nominals and fill roles such as subject, object, or oblique.¹
• Adverbial relativizers (where, when, why, how, etc.) attach as advmod within the RC.
• A possessive pronominal relativizer (whose) modifies a nominal (subject, object, oblique modifier) within the RC the same way as a normal possessive modifier would, i.e., depending on language, it can be nmod:poss, det:poss, nmod or det.
• A relative determiner (which) may modify another noun within the RC via the det relation.

In Enhanced UD (right), the relativizer instead attaches to its antecedent via the ref relation (as the antecedent is directly connected to a role in the RC).

Determinative which cannot be separated from the head on which it is dependent. We take the relativized element in the following example to be the whole nominal, which amount: this is the phrase whose interpretation is given by the antecedent $38,000.²

If the relative clause has a nominal predicate, the relative pronoun may occupy the head position within the clause. In such cases no enhanced relation should be added from its parent to the head noun (because they are the same node). We should only add a nsubj relation from the antecedent to the nsubj of the relative clause (and remove the corresponding nsubj relation between the relative pronoun and the subject).

DZ: This example and explanation is taken from our documentation of enhanced dependencies, where it was added after issue #531. However, the solution appears half-done (why are the remaining children of which not re-attached to chairman?) and it should be elaborated.

A relative clause with no relativizer, like (1), is called a reduced relative clause. One with a relativizer, like (3), is called a nonreduced relative clause.

Relative Markers

In some languages, the relative clause is marked by a morpheme that does not represent any argument within the relative clause; it is thus not a pronoun, and not a relativizer in the sense described above. For example, the de morpheme in Chinese marks the preceding clause as a relative clause. It does not represent the shared argument within the RC; that argument is either realized by zero if it is in a core function, or represented by a personal pronoun (他 tā “he” in the second example; examples from Dixon vol. 2, p. 330³).

• [zh] 你给我邮来的信一半都是英文。 Nǐ gěi wǒ yóu lái de xìn yībàn dōu shì yīngwén. “Half of the letters you send to me are in English.”
• [zh] 我给他送书的人是我弟弟。 Wǒ gěi tā sòng shū de rén shì wǒ dìdì. “The person I gave the book to was my younger brother.”

The relative marker may also be a bound morpheme. In Dyirbal, a relative clause is marked by relative suffix -ŋu replacing the tense suffix on the RC verb (Dixon vol. 2, p. 323).

• [dbl] Jani manigu budilŋaŋu banagañu. “John, who had got the money, returned.” (lit. John-Abs money-Dat get-Antip-Relative-Abs return-Past.)

In Arabic, the relative pronoun may be combined with a resumptive personal pronoun, as in the following example.

• [ar] سيغال التي كانت جدته من مدينة فلاديفوستوك (segal alati kanat jadatuh min madinat fladifustuk) “Seagal, whose grandmother was from Vladivostok”

Notable Properties

In some languages (such as English), relative clauses are typically expected to be finite (people who live in glass houses), although they may also be infinitival (I found a house in which to live; I found a house (for my mother) to live in). Participial clauses (people living in glass houses; students given high marks) are not considered relative clauses in English, they are ordinary adnominal clauses (acl), or even adjectival modifiers (amod) if the participle is tagged ADJ.

Note that the above distinction is not universal, as some languages do not distinguish finite and non-finite verb forms.

An element can be extracted from several levels of embedding:

Semantically, relative clauses may be restrictive (helping to narrow a set of referents), or ascriptive (adding detail about a referent that has already been identified):

• Restrictive
  • I rented the movie which you bought (as opposed to some other one).
• Ascriptive
  • I introduced myself to John, who promptly forgot my name.
  • I rented the movie, which you bought (as opposed to renting).

The restrictive/ascriptive distinction does not affect the UD analysis: all RCs are analyzed with the acl:relcl relation.

The antecedent of an ascriptive relative clause may be even a clause (rather than a nominal):

• I tried to explain myself – which was a bad idea.

DZ: Did we reach a consensus about clausal antecedents? There is the fundamental issue that acl must depend on a nominal. I would prefer something like parataxis:relcl or advcl:relcl over acl:relcl in these cases.

Case-marked Relativizers

If required by the syntactic relations within the RC, the relativizer bears case markers (morphological or adpositional), just as if the position were filled with a full nominal.

• [cs] nebezpečí, kterého se bojíte “danger which you fear” (lit. danger, which.Gen yourself you.fear)
• [cs] nebezpečí, kterému čelíte “danger which you are facing” (lit. danger, which.Dat you.face)
• [cs] nebezpečí, které překonáte “danger which you overcome” (lit. danger, which.Acc you.overcome)
• [cs] nebezpečí, o kterém nevíte “danger which you do not know about” (lit. danger, about which.Loc you.not.know)
• [cs] nebezpečí, se kterým se vyrovnáte “danger which you cope with” (lit. danger, with which.Ins yourself you.cope)

The enhanced relations include deep syntactic relations. Therefore, in case marking languages the enhanced dependencies may link verb dependents that are not in the expected morphological case, required by surface syntax. In the following Czech example, the relative modifier phrase v němž “in which” is obligatorily in the locative case form (Case=Loc). If it were a main clause, the referent dům “house” would have to be in locative too: v domě “in house”. However, here it is in the nominative (Case=Nom), and the enhanced dependency obl going to a nominative dependent is something we would not expect to see, given the morpho-syntactic rules of the language.

Preposition Stranding

In English, a preposition may be left “stranded” in the relative clause (its noun corresponding to the parent of the RC):

• The house (that) you said you wanted to live in is for sale.
  • Non-RC paraphrase: The house is on sale; you said you wanted to live in that house.

The basic UD analysis depends on whether it is a reduced or nonreduced RC. In a nonreduced RC, the relativizer is available to fill a role in the RC, and thus gets marked with the preposition (even if this contributes to the nonprojectivity of the tree):

In a reduced RC, however, the preposition is the only word left from the nominal filling the role in the RC, so the preposition gets promoted to the head of the nominal (in accord with the standard UD treatment of ellipsis).

Relative versus Non-relative Subordinate Clauses

Interrogative Clauses

In languages where the set of relativizers overlaps with interrogative words (such as English or Czech), some clauses may superficially look like relative clauses while they are in fact interrogative clauses (indirect questions). Such clauses are typically complements of verbs of saying or knowing (I don’t know where he lives) but they may also depend on a similar nominal. Compare

• [cs] dům, kde bydlel Hemingway “the house where Hemingway lived”
• [cs] otázka, kde bydlel Hemingway “the question of where Hemingway lived”

In the second example, kde bydlel Hemingway is an adnominal clause (acl) but not a relative clause. There is no coreference between otázka “question” and kde “where”. There is such coreference in the first example, hence the clause is relative there.

Language-specific guidelines must specify whether the wh-word is to be disambiguated between PronType=Int and Rel, or it can be PronType=Int,Rel everywhere and the disambiguation is only at the syntactic level between acl and acl:relcl.

Adverbial Clauses

Pronominal adverbs such as where, when, how, why frequently introduce adverbial clauses (advcl). They can also introduce non-relative adnominal modifier clauses (acl) similarly providing time/place/manner information. In the following example, when is an interrogative/relative adverb (PronType=Int,Rel) but the adnominal clause it introduces is not relative because there is no coreference between headlines and when (but cf. the time when Nixon resigned, here it would be a relative clause).

In general, an adnominal phrase with a wh-adverb is considered a relative clause if the wh-adverb can be paraphrased by in which or similar, or if the head noun reifies the kind of relation (the time when, the place where, the reason why).

Some phrases are ambiguous. The ceremony where/when we became citizens can be interpreted as an RC if the bestowal of citizenship happened during the ceremony (in which interpretation⁴, thus acl:relcl). But the ceremony when we became citizens could be used to refer to a particular ceremony held around the time of becoming a citizen, as opposed to some other ceremony held at some other time; the modifier can be fronted as an advmod within the higher clause: When we became citizens, the ceremony…. This is the acl interpretation.

Interrogative/Relative Adverbs versus Subordinating Conjunctions

Relative adverbs (ADV PronType=Int,Rel) should not be confused with subordinating conjunctions (SCONJ). The former are constituents in the sentence, they refer to circumstances such as location or time of an event. The latter just mark a clause as such a constituent, but they are not constituents themselves. The problem is that some words in some languages seem to oscillate between the two categories. For example, the English word when is clearly an adverb in interrogative contexts: When did he arrive? Like other English question words, it can also be used as a relative adverb: That’s the moment when you’ll realize what really matters to you. In this sentence, when is a constituent of the relative clause and although it is an adverb (and not a nominal), it is coreferential with the nominal argument in the main clause, the moment. However, in other contexts the reference to a particular time period seems to be weakened and the adverb seems to be in the process of grammaticalization towards a subordinating conjunction, somewhat similar to if (which is never an adverb): These companies find it harder to recruit skilled graduates when financial firms can pay higher salaries.

DZ: One option is to say that when is always ADV and never SCONJ, despite possible grammaticalization. Another option is to say that when it introduces an adverbial clause, it is SCONJ, otherwise it is ADV. Adverbial clauses indeed can be introduced by subordinating conjunctions, such as if.

Other languages do not have this problem, as they use distinct words for the adverb and the subordinating conjunction. For example, the possible Czech translations of when are either kdy (ADV) or když (SCONJ):

• [cs] Neodpověděl mi na otázku, kdy hodlá přijet. “He didn’t answer my question of when he was going to come.”
• [cs] V okamžiku, kdy přijede, jsme připraveni mu předvést výsledky. “The moment he arrives, we are ready to show him the results.”
• [cs] Nepřijel, když jsme to nejvíc potřebovali. “He didn’t come when we needed it most.”
• [cs] Když budete mít výsledky, přijedu také. “If you have results, I will come too.”

Free Relatives

Free relatives are nominals containing a relative clause where the relativizer is “fused” with the head of the whole nominal; in other words, there is no explicit modified noun, and its position is filled with a relative pronoun which would otherwise serve as the relativizer (thus these constructions are also known as fused relatives). The relative pronoun is deemed as the head of the construction, thereby receiving a dependency relation reflective of its function in the higher clause, and the rest of the relative clause depends on it as an acl:relcl.

The enhanced representation of a free relative construction will not have a ref relation between the parent nominal and the relativizer because they are the same node. However, a relation between the predicate of the relative clause and the modified nominal will still be added.

DZ: In the above example, we have once again the problem that a relative clause should not be a subtype of acl because it does not modify a nominal. Here it modifies an adverb (where), hence I went with advcl:relcl.

Not all languages have free relatives in the same sense and extent as in English. In Czech, the corresponding construction will typically involve a demonstrative pronoun, which will be modified by a relative clause. The relative clause will have a relative pronoun, which will be case-marked as required within the relative clause, while the case marking of the demonstrative is driven by the requirements of the main clause.

• [cs] Teď se budeme věnovat tomu, čeho se bojíš. “Let’s now attend to what you are afraid of.” (The demonstrative tomu is in the dative, as required by the predicate věnovat se of the main clause. The relative čeho is in the genitive, as required by the predicate bát se of the relative clause. If we swap the functions and make the main clause relative, the forms of the pronouns will change: Teď se bojíš toho, čemu se věnujeme. “You are now afraid of what we are attending to.”)

In rare cases,⁵ the demonstrative can be omitted in Czech. Even when the demonstrative is omitted, the relative pronoun is considered a part of the relative clause, not of the main clause. First, the relative pronoun still follows the case requirements of the relative clause. This is not easy to observe due to limited contexts in which the demonstrative is omitted and to case syncretism, yet it is so. Second, according to the Czech grammar, the relative clause is obligatorily delimited by a comma from the main clause in writing, and the comma is invariably put before the relative pronoun, not after it.

• [cs] Má (to), co si zaslouží. “He has what he deserves.” (main clause accusative, relative clause accusative)
• [cs] Dostanou (to), co jim patří. “They get what belongs to them.” (main clause accusative, relative clause nominative)
• [cs] Lidé ničili (to), čeho se jim dostalo jako božího daru. “People destroyed what they received as a gift from God.” (main clause accusative, relative clause genitive)

DZ: The recommended analysis of the Czech examples is not yet final.

Since the relative pronoun does not belong to the main clause, we cannot analyze the Czech examples the same way as free relatives in English. So what are the options?

1. The “relative” clause is actually an object clause, i.e., the direct object of the main clause is realized as a clause. Consequently, it is attached via ccomp to the predicate of the main clause. This is the approach that is employed in the current conversion of the Czech data. Note that it would be difficult to depart from this approach, as this construction is similar to complement clauses: Vím, co si zaslouží. “I know what he deserves.” The drawback of this approach is that there is no trace of the relativity of the clause, and of the coreference between the argument of the relative/complement clause and the omitted argument of the main clause.
2. There is an elided demonstrative to in the main clause; we use the standard UD treatment of ellipsis. That is, we promote the relative clause to the position of the elided pronoun. Unlike in the previous option, the relation between the predicates of the main and the relative clauses will be obj, not ccomp. There is still the drawback that the construction cannot be recognized as involving a relative clause. An additional drawback is that one has to distinguish between complement clauses (ccomp) and promoted relative clauses (obj), which may be difficult for annotators, and for large corpora such as Czech PDT it is not tractable.
3. The elided demonstrative could be handled similarly to how UD currently handles gapping constructions. The relative clause would be attached to the predicate of the main clause via the obj relation (not orphan, because it is not a link between two orphaned dependents). In the enhanced representation, there would be an empty node representing the demonstrative to, and all the expected relations would be restored as if the demonstrative were overtly present. The objection that we cannot reliably detect these cases in corpora like Czech PDT still holds. The solution would also require a modification of the guidelines regarding the usage of empty nodes (but there is a separate proposal for such an extension). On the other hand, the similarity of the construction to relative clauses would now be recognizable, and so would be the coreference between the missing argument of the main clause and the visible argument of the relative clause. The diagrams below illustrate this last option:

In the Czech examples so far, the elided argument always functioned as a direct object in the main clause. There is another pattern where the common argument acts as a subject both in the main clause and in the relative clause. There are thus no case conflicts, as both positions require the nominative. These examples resemble the English free relatives more closely than the previous ones.⁶

• [cs] Kdo neokrádá stát, okrádá svou rodinu. “He who does not rob the state robs his family.”
• [cs] Kdo kupoval později, může dnes své akcie prodat dráž. “Whoever bought later can sell their shares more expensive today.”
• [cs] Kdo se bojí, nesmí do lesa. “He that is afraid of leaves must not come in a forest.”
• [cs] Co se vleče, neuteče. “All is not lost that is delayed.”

These relative clausal subjects are different from situational clausal subjects in that they represent an entity rather than a situation (event or state). An example of a situational clausal subject would be Že je nevinen, bylo jasně prokázáno. “That he was innocent was clearly proven.” Nevertheless, relative clausal subjects are still clausal subjects, and there is no reason why they could not be attached via the csubj relation. In particular, there is no indication that the relative pronoun belongs to the main clause (rather than to the RC).

☞ TODO: Correlatives in Indian languages.

☞ TODO: Add examples from languages where the fullest representation of the common argument is in the relative clause and there is nothing in the main clause.

Clefts

Relative clauses also occur in the analysis of cleft constructions. In languages that have free relatives, cleft constructions may have them, too.

The Czech equivalent uses a demonstrative pronoun instead of a free relative: Jan je ten, komu chceme pomoct. “Jan is who we want to help.” The demonstrative ten is in the nominative, while the relative komu is in the dative, as required by their respective clauses.

Notes

© 2014–2024 Universal Dependencies contributors. Site powered by Annodoc and brat