Unit 9: Semantic and Syntactic Mismatches

The phenomena covered in this chapter each obscure the surface-to-semantics mapping in a different way:

• Passive — Agent is not the subject; Patient is
• Causatives — a new Causer argument displaces the original subject
• Dative shift — the same participant appears in two different syntactic positions depending on construction
• Raising — the syntactic subject of the matrix clause has no semantic role in that clause
• Control — an implicit argument is bound across clause boundaries
• Expletives — syntactic subject positions are filled by semantically empty elements
• Long-distance dependencies — arguments are displaced from their associated heads
• Argument drop — arguments are omitted from the surface entirely
• Coordinated structures — one argument is shared by multiple predicates

NLP implication: a system that assumes subject = Agent and object = Patient will fail on all of these constructions. Robust semantic interpretation requires construction-aware processing.

Formal description of the English passive:

• The subject of the active (typically the Agent) is demoted to an oblique by-phrase or omitted entirely.
• The direct object of the active (typically the Patient) is promoted to subject position.
• English morphology: be + past participle; Agent (if expressed) marked by by.

Agentless passive: "The window was broken." — The Agent is omitted entirely and cannot be recovered from the sentence alone. This construction is extremely common in formal written text where the agent is unknown, unimportant, or deliberately concealed.

NLP: Passive occurs in approximately 25–35% of clauses in formal written text. SRL systems must identify the Agent even when it is expressed as a by-phrase oblique rather than as the subject.

Anti-passive (common in ergative languages): the Patient is demoted; the Agent remains as subject; the construction signals that the event lacks a specific, affected Patient. Found in Kinyarwanda (Bantu), Chukchi (Siberian), and many other ergative languages.

Impersonal passive: Agent and Patient both demoted; subject position filled by an expletive or left empty. German example:

Middle: "This cloth washes well." The subject is the Patient; the construction expresses a general property of the subject (it is easy to wash), not a specific event with an identified Agent.

NLP: each construction produces a different surface argument pattern for the same underlying event type. SRL systems must handle all three variants.

English ditransitive verbs permit two surface realisations:

• "She gave the book to him." — PP dative: him is Oblique (Goal, inside PP); book is Direct Object (Theme).
• "She gave him the book." — Double-object construction: him is Indirect Object (Goal, bare NP); book is Direct Object (Theme).

Both describe the same event with the same semantic roles. But not all verbs allow both constructions:

• "She explained the theory to him." ✓
• *"She explained him the theory." ✗

NLP: Both constructions should extract the same relation (give; giver; given-thing; recipient), but surface argument positions differ. Relation extraction systems that read surface syntax without normalisation will assign different roles to the same participant.

Causative constructions add a Causer argument as the new subject; the original intransitive subject becomes the object of the causative verb:

• Turkish: gel- (come) → gel-dir- (cause to come / bring)
• Japanese: 食べる (taberu, eat) → 食べさせる (tabe-saseru, cause to eat)

Lexical causative in English: "The doctor grew the bacteria" — a causative use of an otherwise intransitive verb. NLP: causative constructions add an argument and shift the position of the original subject; SRL must detect this alternation and correctly assign the Causer role.

Expletives, pleonastic pronouns, and dummy auxiliaries fill syntactic positions that have no semantic content:

• "It is easy to leave." — it is semantically empty; to leave is the logical subject.
• "There are three solutions." — there is empty; three solutions is the logical subject.
• French: "Il pleut" (it-rains) — expletive subject.

Mandarin Chinese has relatively few such constructions, but some discourse-initial particles serve a similar structural function.

NLP: Empty elements occupy syntactic argument positions and must be identified as non-referential to avoid incorrect entity extraction. Coreference resolution systems must not link expletive it or there to prior entities in the discourse.

Key properties of expletives:

• Syntactically present, semantically empty
• Cannot be questioned: *"What rained?"
• Cannot be replaced by a referential pronoun: *"It rained and so did it."

Examples: "It rained." / "It seems that she left." / "There is a problem." / "There appear to be errors."

NLP applications affected by expletive detection:

1. Coreference resolution — expletive it must not be linked to a prior referent.
2. SRL — the expletive position should not receive a semantic role label.
3. Machine translation — many languages do not use expletives; the logical subject must be identified and translated appropriately.

"She seems to know the answer."

Seem does not assign a semantic role to its syntactic subject. She is raised from the embedded clause (she = subject of know). Evidence for raising:

• "It seems that she knows the answer" — same semantics; the expletive it can freely replace she as the syntactic subject of seem, confirming that seem does not require a meaningful subject.
• "There seem to be problems" — the semantically empty there raises into subject of seem, showing that any NP, even an expletive, can occupy this position.

NLP: raising verbs create a mismatch — the syntactic subject of the matrix clause is actually the semantic argument of the embedded predicate. SRL systems must detect this and assign the semantic role to the embedded predicate, not the matrix one.

"She tried to leave." — try is a subject-control verb.

• She is the syntactic and semantic subject of tried (she is performing the trying).
• She is also the understood subject of leave (she is the one who would leave).
• The two subjects are coreferential.

Contrast with raising: with a control verb, the matrix subject has a semantic role with respect to the matrix predicate (try predicts that its subject is attempting something). Diagnostic: *"It tried that she left" is ungrammatical — try requires a meaningful subject.

Object control: "She persuaded him to leave" — him is the object of persuaded AND the understood subject of leave.

NLP: control relations create implicit argument bindings — the understood subject of the embedded predicate must be recovered for full semantic interpretation. This is distinct from explicit coreference; it is licensed by the lexical properties of the control verb.

In some languages, multiple predicates combine to license a single argument structure:

• Japanese compound verbs: two verbs fuse into a unit. 食べ始める (tabehajimeru, "start eating") — the arguments of the whole complex are distributed across both verbal elements.
• Urdu/Hindi light verb constructions: a noun or adjective combines with a light verb (do, give, take) to express an event: kām karnā (work do-INF = "to work"). The semantic content is in the noun; the argument licensing is shared.

NLP: argument extraction from complex predicate constructions requires joint processing of multiple predicates. Sequential independent extraction of one predicate at a time will miss the full argument structure.

Coordination creates non-standard dependency patterns:

• One-to-many: "John and Mary left." — John is subject of left AND Mary is subject of left; two subjects share one verb.
• Many-to-one: "She sang and danced." — she is subject of both sang and danced; one subject is shared across two verbs.
• Gapping: "She visited Paris, and he, London." — the verb visited is omitted from the second conjunct; its presence must be inferred.

NLP: coordination creates non-projective dependency structures that simple projective parsers cannot handle. Gapping and gap-coordination require dedicated parser handling; many state-of-the-art parsers still struggle with these constructions.

In questions, relative clauses, and topicalisations, an argument is displaced from its canonical position to a position earlier in the sentence — potentially far away:

• Wh-question: "What₁ did the dog chase t₁?" — what is the object of chase but appears sentence-initially.
• Relative clause: "The cat [that the dog chased t]" — gap inside relative clause.
• Topicalisation: "This book₁ I read t₁ last year."
• Deeply embedded: "What₁ did she say [that she believed [that he thought [that she had done t₁]]]?"

NLP: long-distance dependencies require unbounded dependency grammars or equivalent mechanisms. Transformer attention can in principle cover long-distance relations, but LLMs still underperform on complex nested long-distance dependencies, losing track of the filler by the time they reach the deep gap position.

Most languages keep modifiers adjacent to their head noun. Some allow or require separation:

• German extraposition: a relative clause may be moved to the end of the main clause even when its head noun is earlier — "Ein Mann ist gekommen [der niemand kannte]" (A man has come who nobody knew) — the relative clause is extraposed away from Mann.
• Dutch: PP modifiers of NPs can appear clause-finally, separated from their head.
• English extraposition: "A man arrived [who nobody knew]" — relative clause is separated from head noun man.
• Japanese: relative clauses are pre-nominal and can be very long; the entire clause precedes the noun it modifies.

NLP: parsers that assume adjacency for modifier attachment will fail on these constructions. Long-distance modifier attachment is a known source of parser error.

Languages vary considerably in how freely arguments can be omitted:

• Japanese: subjects and objects are freely dropped when contextually recoverable. 行った (itta, "went") — who went? Recoverable from context.
• Spanish pro-drop: subject pronouns are omitted because verb agreement encodes person and number. "Habla bien" (speaks well = "he/she speaks well").
• English object drop: "She was eating" — implicit object; the thing being eaten is unspecified.
• Italian: "Piove" (rains = "it is raining") — impersonal construction without an expletive subject.

NLP: argument drop requires systems to recover the referent of the dropped argument. This is zero-anaphora resolution, a major challenge in Japanese, Chinese, and pro-drop languages generally — even state-of-the-art models perform significantly below human levels on these tasks.

Not all dropped arguments resolve to a specific contextual referent:

• Definite drop (contextually recoverable): in Japanese subject/object drop, the referent is typically definite — the hearer knows who is meant from the preceding discourse. Zero-anaphora resolution aims to identify this antecedent.
• Indefinite drop (generic/unknown agent): impersonal constructions use a generic dropped subject with no specific referent. French on, German man, English one: "One should always be careful" — the subject is indefinite; there is no antecedent to resolve.

NLP: distinguishing definite drop (resolve to a specific antecedent) from indefinite drop (treat as generic/unknown) affects both coreference resolution and semantic role assignment. A coreference system that attempts to resolve every dropped argument to a prior entity will fail on impersonal indefinite constructions.