Ellipsis (Linguistic)

How ellipsis works

Open any ordinary conversation transcript and you will find sentences that are, on the surface, ungrammatical. "Yes, I can." "No, she didn't." "A coffee, please." "The big one." Each is a fragment, but each communicates a full proposition because the listener fills in unpronounced material from the surrounding discourse. Linguistic ellipsis is the formal study of when this filling-in is licensed and how the missing structure relates to the visible structure.

The central question is what occupies the silent slot. Three positions dominate the literature:

1. PF-deletion. Full syntactic structure is built and interpreted; phonology then fails to pronounce a designated subtree. Sag (1976) introduced the position; Merchant (2001) gave it the modern formulation.
2. LF-copying. The ellipsis site contains an empty pro-form that is filled at logical form by copying semantic structure from the antecedent. Williams (1977), Hardt (1993).
3. Pro-form / null anaphor. The ellipsis site is a designated empty category — a null counterpart to do so or it — interpreted directly without internal syntax. Lobeck (1995).

Each view captures different data. Movement out of an elided VP ("I know which book John read but not which one Mary did") argues for PF-deletion: the wh-phrase escapes from internal structure that must be syntactically real. Discourse-only antecedents (a doorbell rings; you say "I'll get it" — no linguistic antecedent for "it" exists) argue for semantic recovery. Modern hybrid theories (Elbourne 2008) accept that different elliptical constructions sit at different points on the syntax-semantics gradient.

The major types of ellipsis

Type	Example	What is elided	Licensor / remnant
VP-ellipsis (VPE)	John can swim, and Mary can too.	Verb phrase after auxiliary	Auxiliary, modal, infinitival to, do-support
Sluicing	John ate something, but I don't know what.	Embedded IP/CP under a wh-remnant	Wh-phrase in Spec,CP
Gapping	John ate fish, Mary chicken.	Finite verb (and sometimes more) in a coordinate clause	Coordination; two contrastive remnants
Pseudogapping	John gave Bill a book, and Mary will Sue.	Verb only; auxiliary and object remain	Auxiliary plus contrastive XP remnant
NP-ellipsis	John's car is faster than Mary's.	Head noun and modifiers after a possessor or determiner	Possessive marker or strong determiner
Stripping (bare argument ellipsis)	John left, and Mary too.	Everything but a single XP and a polarity element	Coordinator plus too, also, not
Fragment answer	Q: What did Mary eat? — A: Fish.	Everything but the focused constituent	Question-answer pair; focus on remnant
Antecedent-contained deletion	John read every book that Mary did.	VP whose antecedent properly contains the ellipsis site	Quantifier raising of object out of VP

The diagnostics distinguishing these are subtle. Pseudogapping looks like VPE but leaves an object remnant; gapping refuses to nest under embedding ("*John said that he ate fish, and Bill that chicken" is bad). Sluicing tolerates island violations that overt wh-movement would not — the "island repair" effect (Ross 1969; Merchant 2008). NP-ellipsis is licensed in English only by certain determiners; possessor 's licenses it ("Mary's"), demonstrative "this" does not ("*this" alone). These language-specific licensing facts are part of why the typology runs deep.

Worked example: sluicing in detail

Take the sentence John ate something, but I don't know what. The second clause looks like an isolated wh-word, but it is interpreted as "I don't know what John ate." Merchant's (2001) PF-deletion analysis works as follows:

1. The embedded clause is built fully: [_CP what [_C' [_IP John ate t]]].
2. The wh-phrase "what" moves to Spec,CP as in any embedded question.
3. The IP "John ate t" satisfies the e-GIVENness identity condition relative to the antecedent IP "John ate something" (modulo focus on the wh-existential pair).
4. An [E] feature on C licenses non-pronunciation of its IP complement at PF.
5. The result: a single remnant "what" is heard, but the meaning is the full embedded question.

This account predicts that sluicing should support overt connectivity effects — case marking, preposition stranding, complementizer choice — that match the elided structure rather than the surface form. Russian sluiced wh-phrases bear the case the elided verb assigns; German sluicing reflects whether the antecedent verb takes accusative or dative. These "form-identity" effects (Merchant 2008) are central evidence for syntactic structure inside the silent IP.

Counterarguments and alternative views

Not all syntacticians accept that ellipsis sites contain full hidden syntax. Three classes of counterargument shape the modern debate:

• Sag-style mismatches. Voice mismatches in pseudogapping ("The information was leaked by Mary, just as John had hoped to") and binding-condition mismatches show that some elliptical relations are looser than full syntactic identity would predict. Critics argue this favors a semantic identity condition.
• Construction grammar. Goldberg and others treat fragments and elliptical replies as constructions in their own right — pairings of form and meaning learned independently — rather than derivatives of full clauses. This sidesteps the problem of recovering hidden structure by denying it exists.
• Discourse-driven views. Hardt (1993, 1999) and Hankamer-Sag (1976) distinguish "surface anaphora" (requiring syntactic antecedent, e.g. VPE) from "deep anaphora" (allowing discourse antecedents, e.g. do it). They argue ellipsis straddles both — discourse alone licenses some null sites, undermining a uniformly syntactic theory.

The debate is unresolved. Most working syntacticians treat PF-deletion as the default for VPE, sluicing, gapping, and fragments, while reserving semantic or pro-form analyses for cases that resist syntactic identity.

Variants and edge cases

• Antecedent-contained deletion. "John read every book that Mary did." Solved by quantifier raising the object at LF (May 1985; Sag 1976) so the antecedent VP shrinks below the ellipsis site.
• Comparative ellipsis. "Mary read more books than John did." A close cousin of ACD; Kennedy (1997) extends the QR analysis.
• Backward ellipsis. "Whoever can ___ should bring the dessert" in cleft-like structures with cataphoric resolution. Rare in English, more common in head-final languages.
• Right-node raising. "John likes, and Mary hates, garlic." Often grouped with ellipsis but typically analyzed as across-the-board movement (Postal 1974).
• Argument ellipsis (Japanese). A bare null object that licenses sloppy readings. Otani and Whitman (1991) argued this is a covert VPE; Oku (1998) and Saito (2007) argued for a separate argument-ellipsis mechanism.
• Multiple sluicing. "John gave something to someone, but I don't know what to whom" — multiple wh-remnants, attested in Bulgarian, Romanian, and Japanese.

Common pitfalls in analysis

• Confusing ellipsis with pro-forms. "John left, and Bill did too" (VPE) versus "John left, and Bill did so too" (do-so anaphora). The latter has internal pro-form structure; the former has hidden full structure. Diagnostics differ — extraction is allowed under VPE but not under do so.
• Treating gapping as VPE. Gapping requires coordination and contrastive remnants; VPE does not. Gapping cannot embed; VPE can.
• Ignoring focus structure. Sloppy readings, fragment answers, and contrastive remnants are governed by focus, not bare syntax. Rooth's (1992) focus identity and Merchant's e-GIVENness both encode this.
• Forgetting recoverability. A deletion that is not recoverable from context is not ellipsis — it's ungrammatical or a fragment that fails to communicate. The Aspects (1965) recoverability principle remains baseline.
• Overgeneralizing from English. English VPE is unusually well-licensed because of its rich auxiliary system. Many languages license different cuts of the structure; do not assume English is universal.