Object Relative Clauses

• Object relative clauses (ORCs) are relative clauses where the head noun surfaces are the object of the embedded verb within these structures. For example in the mandarin text 'the tycoon invite [GAP] de official.' ORCs require more syntactic work that involves more semantic and Structural dependencies
• Key psychological processes driving OCRs comprehension include retrieving the object dependency (these are more complex by nature) and the comprehension time, which tends to be longer for ORCs than for SRCs across many different languages.
• Neural language models, such as BERT, can handle ORCs but show limitations in reliably converting semantic information into structural analyses or maintaining animacy and grammaticality in predictions.

Object relative clauses (ORCs) are relative clauses in which the head noun is interpreted as the object of the embedded verb, while an overt or null subject occupies the embedded subject position. In English, the books that my cousin likes contains an object gap after likes: the head noun books is interpreted as the object of likes. ORCs contrast with subject relative clauses (SRCs), such as the children who eat vegetables, in which the head noun corresponds to the embedded subject. ORCs are studied in formal semantics, categorical and vector-based composition, psycholinguistics, language acquisition, parsing, and neural language-model evaluation.

1. Grammatical structure and cross-linguistic realization

In an ORC, the antecedent or head noun is interpreted as filling a missing object position inside the relative clause. The embedded subject is overt in examples such as the dress that I brought yesterday and the men whom Mary loves. A schematic representation is:

$$[\text{NP the dress}]_i\ [\text{RC that } I\ brought\ \_\_i\ yesterday]$$

The gap is linked to the antecedent by a dependency. In a subject relative, by contrast, the gap occupies the embedded subject position: the woman who sought to attract stares. This distinction determines the thematic-role relation between the antecedent and the embedded predicate.

English marks the contrast through word order and, in some cases, relativizer form. In men who hate women, the antecedent is the subject of hate; in men whom women hate, it is the object. American English permits omission of the relativizer in restrictive ORCs:

• the letter which she hands to Kevin;
• the letter that she hands to Kevin;
• the letter she hands to Kevin.

Relativizer omission is not normally licensed in the corresponding subject-relative configuration: a woman who sought to attract stares contrasts with the unacceptable a woman sought to attract stares (Mosbach et al., 2020).

Chinese relative clauses are prenominal and end in the relativizer de. In a subject relative, the gap is the embedded subject:

$$[\text{GAP}_i\ \text{invite tycoon de}]\ \text{official}_i$$

In an object relative, the gap is the embedded object:

$$[\text{tycoon invite GAP}_i\ \text{de}]\ \text{official}_i$$

The subject-relative dependency is longer than the object-relative dependency when distance is measured as the number of words between the gap and the head noun. This configuration reverses a familiar English pattern in which ORCs are often more difficult than SRCs (Vasishth et al., 2017).

Dutch illustrates a different source of ambiguity. Because subordinate clauses have essentially subject–object–verb order, mannen die vrouwen haten can mean either men who hate women or men whom women hate. The surface string does not determine whether the gap associated with die is the subject or object of haten. The ambiguity is therefore derivational: two syntactic proofs connect the same relative pronoun to different verb arguments (Moortgat et al., 2017).

Mandarin also permits complex recursive configurations involving subject- and object-gapped relatives. The four combinations studied in child production are:

Type	Configuration
SORRC	Subject-gapped RC embedded in object-gapped RC
OORRC	Object-gapped RC embedded in object-gapped RC
OSRRC	Object-gapped RC embedded in subject-gapped RC
SSRRC	Subject-gapped RC embedded in subject-gapped RC

These configurations show that object-gap status alone does not determine difficulty: OORRCs were produced earliest, whereas SSRRCs were the most difficult overall (Yang et al., 2024).

2. Categorial and vector-based semantics

A compositional vector-based analysis treats the semantic contribution of an ORC as structural rather than as a context-independent corpus vector for whom, that, or which. The relative pronoun connects the head noun to a position inside the embedded clause, passes information between the clause and the noun phrase, combines compatible noun information, and discards the sentence or truth-value component after modification (Sadrzadeh et al., 2014).

In a pregroup grammar, the basic types are noun phrase $n$, sentence $s$, and their adjoints. A transitive verb has type:

$n^r\otimes s\otimes n^l$

The subject and object relative pronouns receive distinct types:

$n^r\,n\,s^l\,n$

for a subject relative, and

$n^r\,n\,n^{ll}\,s^l$

for an object relative. The double left adjoint $n^{ll}$ records the movement of the object from the end of the clause to the beginning of the relative construction. Under a strongly monoidal functor into finite-dimensional vector spaces,

$F(n)=N,\qquad F(s)=S,$

and compactness identifies $$[\text{GAP}_i\ \text{invite tycoon de}]\ \text{official}_i$$0 with $$[\text{GAP}_i\ \text{invite tycoon de}]\ \text{official}_i$$1. The corresponding semantic types are:

$$[\text{GAP}_i\ \text{invite tycoon de}]\ \text{official}_i$$2

for the subject-relative pronoun and

$$[\text{GAP}_i\ \text{invite tycoon de}]\ \text{official}_i$$3

for the object-relative pronoun. The position of the noun and sentence wires determines which argument is contracted and which noun dimension is combined with the head noun.

The categorical machinery uses compact closed categories with tensor products, evaluation maps $$[\text{GAP}_i\ \text{invite tycoon de}]\ \text{official}_i$$4, and coevaluation maps $$[\text{GAP}_i\ \text{invite tycoon de}]\ \text{official}_i$$5. Evaluation performs ordinary grammatical application by contracting a verb with its arguments. Coevaluation introduces identity-like wires that bridge the relative clause and the head noun. The analysis distinguishes grammatical flow, represented by $$[\text{GAP}_i\ \text{invite tycoon de}]\ \text{official}_i$$6-maps, from pronominal or semantic flow, represented by $$[\text{GAP}_i\ \text{invite tycoon de}]\ \text{official}_i$$7-maps.

Frobenius algebras provide the operations required for copying, combining, and discarding information. For a basis $$[\text{GAP}_i\ \text{invite tycoon de}]\ \text{official}_i$$8, the standard structure is:

$$[\text{GAP}_i\ \text{invite tycoon de}]\ \text{official}_i$$9

$$[\text{tycoon invite GAP}_i\ \text{de}]\ \text{official}_i$$0

$$[\text{tycoon invite GAP}_i\ \text{de}]\ \text{official}_i$$1

and

$$[\text{tycoon invite GAP}_i\ \text{de}]\ \text{official}_i$$2

The comultiplication $$[\text{tycoon invite GAP}_i\ \text{de}]\ \text{official}_i$$3 copies basis-indexed information, while multiplication $$[\text{tycoon invite GAP}_i\ \text{de}]\ \text{official}_i$$4 retains only matching basis indices. For vectors, multiplication is componentwise:

$$[\text{tycoon invite GAP}_i\ \text{de}]\ \text{official}_i$$5

The counit $$[\text{tycoon invite GAP}_i\ \text{de}]\ \text{official}_i$$6 discards the sentence or truth-value component; the unit $$[\text{tycoon invite GAP}_i\ \text{de}]\ \text{official}_i$$7 supplies an all-basis-elements sentence component. Commutativity and speciality permit connected diagrams to be normalized as spiders, so multiple compatible Frobenius operations can fuse.

The formal meanings of the relative pronouns are:

$$[\text{tycoon invite GAP}_i\ \text{de}]\ \text{official}_i$$8

for the subject relative, and

$$[\text{tycoon invite GAP}_i\ \text{de}]\ \text{official}_i$$9

for the object relative. Their difference lies in the location of the sentence and identity wires. After grammatical contraction, the subject-relative construction combines the embedded subject information with the head noun, while the object-relative construction combines the embedded object information with the head noun.

In a truth-theoretic instantiation, a transitive verb is represented as a weighted relation:

$n$0

For an object relative, the resulting vector is:

$n$1

The output is therefore a weighted sum of candidate objects. In the Boolean case, men whom Mary loves contains exactly those men $n$2 for which $n$3 belongs to the love relation.

In a corpus-statistical instantiation, noun vectors contain corpus-derived coefficients and verbs are matrices or tensors. The object-relative composition is:

$n$4

The transpose expresses reversal of argument direction: the subject vector is applied to the verb to extract candidate objects, which are then multiplied componentwise with the head-noun vector. For men whom Mary loves, the semantic operation is therefore an intersection-like reweighting of the men vector by the objects selected by Mary loves.

The Dutch analysis uses one lexical recipe for die:

$n$5

The same tensor is used for both subject and object readings. Different syntactic proofs generate different index-linking patterns: in one proof the extracted argument contracts with the verb’s subject index; in the other it contracts with the object index. Syntax determines argument linking, while the Frobenius tensor supplies copying, diagonal matching, sentence summation, and intersective combination (Moortgat et al., 2017).

3. Processing and comprehension

Chinese ORCs provide a test of two theories of dependency completion. A distance-based account predicts that longer dependencies produce uniformly longer retrieval times. A direct-access account treats reading times as a mixture of relatively fast successful retrievals and slower trials involving initial retrieval failure followed by reanalysis.

The relevant Chinese constructions are prenominal. In the subject relative the official who invited the tycoon, the gap is the subject of invite. In the object relative the official who the tycoon invited, the gap is the object. The subject-relative dependency is longer under a word-intervening measure, so a simple distance account predicts an object-relative advantage at the head noun.

Raw reading times initially appeared to support this prediction, with an estimated object-relative advantage of approximately $n$6 ms:

$n$7

On log-transformed reading times, however, the estimate was smaller and uncertain:

$n$8

Subject-relative times were more dispersed and contained more extreme values. This distributional asymmetry motivated a hierarchical mixture model in which the slow component represents retrieval failure and reanalysis. The model used log-normal distributions, with condition-specific failure probabilities $n$9 and $s$0, a displacement parameter $s$1, and separate variances for successful and failed retrievals.

For the original data, the estimated failure probabilities were:

$s$2

Their difference was $s$3, with a 95% credible interval of $s$4 and an approximately $s$5 posterior probability that the difference was positive. The failure/reanalysis component was substantially more variable:

$s$6

compared with

$s$7

for successful retrievals. The replication produced the same qualitative pattern:

$s$8

with a posterior probability of a positive difference of approximately $s$9 in the main results discussion.

Ten-fold cross-validation favored the mixture model. In the original data, the expected log predictive density was approximately $n^r\otimes s\otimes n^l$0 for the mixture model and $n^r\otimes s\otimes n^l$1 for the hierarchical model, a difference of approximately $n^r\otimes s\otimes n^l$2 in favor of the mixture model. In the replication, the corresponding values were approximately $n^r\otimes s\otimes n^l$3 and $n^r\otimes s\otimes n^l$4, with a difference of approximately $n^r\otimes s\otimes n^l$5.

The results support a probabilistic interpretation of ORC comprehension. Object relatives appear easier in these Chinese data primarily because they generate fewer failed retrievals and fewer extreme reading times, not because every successful ORC dependency is completed uniformly faster. Distance may affect the probability of failure without deterministically increasing the duration of every successful retrieval. The mixture components are latent, however, and reading times alone do not establish that they correspond exactly to successful retrieval and reanalysis.

German corpus work addresses a different issue: whether relative clauses remain adjacent to their head noun or are extraposed. It does not isolate ORCs. In-situ relatives had fewer discourse-new referents and lower average surprisal than extraposed relatives. The study therefore suggests that relative clauses with more accessible and predictable content are more likely to remain integrated, whereas less predictable or discourse-new material is more likely to be outsourced to the post-field. Applying this result specifically to ORCs remains an extrapolation because the corpus did not classify relativized grammatical function (Speyer et al., 2017).

4. Acquisition and recursive object relatives

Mandarin recursive relative clauses show how object-gap dependencies interact with semantic reversibility during acquisition. The study examined 429 Mandarin-speaking children aged 3–11 and 80 adults using a $n^r\otimes s\otimes n^l$6 design crossing four recursive syntactic types with two semantic conditions.

In irreversible internal semantics (IIS), noun–verb combinations strongly constrain thematic roles, as in the sister eats the banana. In reversible internal semantics (RIS), the nouns are mutually compatible with the relevant predicates, so lexical semantics does not determine the roles. Every item was presented with irreversible external semantics (IES): a speech–visual scene specified one event configuration.

Children produced recursive relatives more accurately in IIS–IES than in RIS–IES conditions. Significant effects were found for age, semantic condition, syntactic type, and the Age × Condition × Syntax interaction:

$n^r\otimes s\otimes n^l$7

$n^r\otimes s\otimes n^l$8

$n^r\otimes s\otimes n^l$9

The proposed developmental path has two stages. Initially, syntax is acquired through an interface with irreversible internal semantics:

$n^r\,n\,s^l\,n$0

Later, children can use external visual or discourse semantics even when lexical relations are reversible:

$n^r\,n\,s^l\,n$1

The developmental ordering was:

$n^r\,n\,s^l\,n$2

where greater-than indicates greater structural complexity and later acquisition. OORRCs were acquired at approximately six years in IIS–IES and seven years in RIS–IES. SORRCs were acquired at approximately six and eight years, OSRRCs at seven and nine years, and SSRRCs at eight and ten years. Thus, SORRCs, OSRRCs, and SSRRCs showed an approximately two-year semantic delay, whereas OORRCs showed a one-year delay.

These findings do not support the claim that object-gapped relatives are uniformly more difficult than subject-gapped relatives. OORRCs contained two object gaps but were the earliest and most accurate configuration; SSRRCs contained only subject gaps but were the latest and least accurate. Difficulty was better associated with recursive dependency configuration, dependency distance, noun interruptions, and competing structural analyses.

Younger children commonly simplified the targets into single nouns, simple declaratives, coordinated clauses, or one of the embedded relatives. Older children increasingly produced complex but non-target structures, including declaratives containing relative clauses and relatives with conjoined or serial verb phrases. The error pattern supports a developmental transition from simpler structures toward increasingly hierarchical representations rather than a binary transition from absent to adult syntax.

5. Parsing ambiguity and neural LLMs

Dutch ORCs constitute a strong test of whether parsers use semantic and discourse information to resolve structural ambiguity. In de dokter die de patiënt geneest, the head noun can be the subject of geneest—“the doctor who cures the patient”—or its object—“the doctor whom the patient cures.” The ambiguity is particularly difficult in reversible or weak-preference subject–verb–object triples.

A study using Lassy Large extracted 3,304 irreversible triples, 370 triples with a strong preference for the original subject–object order, and 724 weak-preference reversible triples. Two BERT-based parsing regimes were compared: a neurosymbolic neural proof-net parser and a Universal Dependencies (UD) parser (Wijnholds et al., 2023).

The proof-net parser represents the ambiguity in the formula assigned to the relative pronoun. A subject-relative analysis uses a subject-gap dependency such as

$n^r\,n\,s^l\,n$3

whereas an object-relative analysis uses

$n^r\,n\,s^l\,n$4

The UD parser expresses the contrast through atomic token labels such as:

$n^r\,n\,s^l\,n$5

and

$n^r\,n\,s^l\,n$6

Without a preceding sentence, both systems showed a strong subject-relative bias. For irreversible reversed triples, the proof-net parser assigned an object-relative analysis in $n^r\,n\,s^l\,n$7 of cases, compared with $n^r\,n\,s^l\,n$8 for the UD parser. For strong-preference reversed triples, the corresponding figures were $n^r\,n\,s^l\,n$9 and $n^r\,n\,n^{ll}\,s^l$0; for weak-preference reversed triples, $n^r\,n\,n^{ll}\,s^l$1 and $n^r\,n\,n^{ll}\,s^l$2.

A preceding sentence improved object-relative interpretation, particularly in irreversible cases. For the proof-net parser, object-relative accuracy in irreversible reversed conditions increased to $n^r\,n\,n^{ll}\,s^l$3; for the UD parser it reached $n^r\,n\,n^{ll}\,s^l$4. The improvement was much smaller for reversible cases, where the subject-relative bias remained strong.

Task-specific fine-tuning corrected the bias more effectively in the proof-net parser. Its object-relative accuracy reached $n^r\,n\,n^{ll}\,s^l$5 for irreversible cases and approximately $n^r\,n\,n^{ll}\,s^l$6–$n^r\,n\,n^{ll}\,s^l$7 in several reversible conditions. The UD parser also responded to contextual manipulation, but retained a stronger asymmetry: object-relative accuracy was $n^r\,n\,n^{ll}\,s^l$8 and $n^r\,n\,n^{ll}\,s^l$9 in regular-prior strong- and weak-preference conditions, respectively.

The results suggest that contextual language-model representations can contain information relevant to object-relative disambiguation without the attached parser reliably converting that information into a structural analysis. The proof-net architecture was more responsive to bias correction, plausibly because it represents the head–gap relation and subject/object gap type compositionally. This adaptation incurred costs: ordinary parsing performance declined, and jointly fine-tuning BERT and the parser led to lower general parsing quality than fine-tuning the parser alone.

Masked LLMs also show a distinction between recoverable grammatical information and robust syntactic-semantic computation. A study of BERT, RoBERTa, and ALBERT used naturalistic American English data, diagnostic cases, masked relativizer prediction, and masked antecedent prediction (Mosbach et al., 2020). Probing classifiers achieved overall accuracies of $n^{ll}$0 for BERT, $n^{ll}$1 for RoBERTa, and $n^{ll}$2 for ALBERT. In object-relative omission cases, however, the models had to distinguish grammatical omission from ungrammatical subject-relative omission.

For masked relativizer prediction in ORCs, the models performed best on that and less reliably on which. MP@1 for object who, which, and that was respectively $n^{ll}$3 for BERT; $n^{ll}$4 for RoBERTa; and $n^{ll}$5 for ALBERT. Qualitative judgments showed that grammaticality and animacy were generally better captured than plausibility or exact target recovery. Models could select a grammatically possible relativizer without recovering the corpus form or a semantically appropriate antecedent.

Masked antecedent prediction was more difficult. Object-relative MP@1 for who, which, and that was $n^{ll}$6 for BERT; $n^{ll}$7 for RoBERTa; and $n^{ll}$8 for ALBERT. The models often preserved animacy or grammaticality while selecting overly general, unrelated, or semantically implausible nouns. These results show that high probing accuracy does not establish reliable antecedent identification, thematic-role assignment, case marking, or plausibility evaluation.

A causal intervention study examined whether BERT uses relative-clause boundary information in agreement prediction. In the skater that the officers love [MASK] happy, the matrix copula must agree with skater, not the plural embedded subject officers. AlterRep intervenes on a linearly identified RC subspace while preserving the remaining representation (Ravfogel et al., 2021).

Positive interventions that made the masked copula appear to be inside an RC increased the probability of the incorrect plural form. For BERT-base, effects were concentrated in middle layers, especially layers 5–8 of 12; the positive intervention increased error probability by as much as 14 percentage points. On initially incorrect cases, the negative intervention, which made the copula appear outside the RC, reduced error probability by as much as approximately 16 percentage points.

The effects generalized partially across unreduced and reduced ORCs, SRCs, and passive relatives, although same-type effects were larger than cross-type effects. Random-subspace interventions did not produce the same interpretable pattern. The results support the existence of RC-related representations that are causally involved in agreement prediction, but they do not establish that BERT implements an explicit symbolic parse tree or a complete theory of ORCs.

6. Attachment ambiguity and multilingual computational behavior

Relative-clause attachment ambiguity differs from object relativization. In the cover of the song that features the famous violinist, the RC may attach high to DP1 (the cover) or low to DP2 (the song). The entire complex DP can occur in object position, as in the cameraman hated the relative of the actor who drank. The RC in such examples is generally a subject relative—who drank—rather than an object relative. The relevant manipulation concerns whether the RC modifies the higher or locally adjacent noun, not whether the gap is an embedded object (Lee et al., 4 Mar 2025).

A multilingual study evaluated Claude 3 Opus, Gemini-1.5 Pro, GPT-3.5, GPT-4o, and Llama 3 70B in English, Spanish, French, German, Japanese, and Korean. It varied RC length and the syntactic position of the complex DP. English, Spanish, French, and German were treated as head-initial languages; Japanese and Korean as head-final languages.

English showed a strong low-attachment preference across models: high-attachment responses ranged from $n^{ll}$9 for GPT-4o to $F(n)=N,\qquad F(s)=S,$0 for GPT-3.5. This corresponded to the reported human English preference for low attachment. Spanish, French, and German produced substantial model variation. Japanese was predominantly high-attachment oriented in the model results, while Korean showed extreme low-attachment preferences despite the cited human high-attachment preference.

The effective data were strongly language-dependent. The overall unusable or excluded proportion was $F(n)=N,\qquad F(s)=S,$1, but it was $F(n)=N,\qquad F(s)=S,$2 for Japanese and $F(n)=N,\qquad F(s)=S,$3 for Korean. Models often responded in English to Japanese and Korean prompts, suggesting translation-mediated or English-centered processing. The study therefore treats apparent competence in these languages cautiously.

Length and position effects were not uniform. English showed no significant length, position, or interaction effect. Spanish, French, and German showed significant position contrasts but no significant length or interaction effects. Japanese showed a significant length effect but no significant position or interaction effect. Korean showed no significant predictor, consistent with its high exclusion rate.

The proposed mechanisms include locality, default low attachment, the Balanced Sister hypothesis, prosodic weight, focus, language-specific attachment preferences, translation-mediated processing, resource imbalance, and model-specific training differences. These findings bear on ORCs only indirectly: they concern RC attachment to complex DPs in object position rather than a factorial comparison of subject and object relatives. They nevertheless demonstrate that object-position structure, constituent length, language-specific word order, and model architecture can affect the interpretation of relative-clause dependencies.

Across formal, cognitive, developmental, and computational research, ORCs are consequently not characterized by a single source of difficulty. Their interpretation depends on the grammatical position of the gap, the direction of argument composition, dependency length, retrieval failure probability, recursive embedding, lexical thematic reversibility, discourse grounding, corpus frequency, attachment structure, and the representational architecture of the parser or LLM. Structural analyses distinguish ORCs from SRCs; semantic models encode the head–object relation compositionally; processing models treat difficulty probabilistically; acquisition studies show developmental reweighting between internal and external semantics; and neural-model studies distinguish the presence of decodable information from its causal use in prediction.