Elliptical Feature Agreement Seungwan Ha Korea University

1.

Elliptical Feature

Ellipsis is a common redundancy-avoidance strategy we adopt in communications. The speaker would often execute unpronounciation of a phrase when its antecedent is found in the previous discourse. For example, VP and TP can be elided in (1-2) respectively since their semantically identical antecedent is posited in the first conjunct. Note that as far as grammaticality goes, there would be nothing wrong if the phrases in the second conjunct were pronounced, but ellipsis contributes to improvement of discourse coherence. In (1), the elided VP love studying syntax is recovered by the VP in the first conjunct. Assuming that wh-movement takes place in the second conjunct and then TP is elided (Merchant 2001), the TP can be recovered by the matrix TP Amanda has to pick t in (2). (1)

Joey loves studying syntax, and Mary does <>, too.

(2)

Amanda has to pick one contestant, but she hasn’t decided who <>.

How to calculate “identity” between the antecedent and elided phrases has been elaborated by Merchant’s (2001, 2004, 2008a) e-GIVENness analysis under the minimalist framework (cf. Lobeck 1995). Merchant’s analysis of ellipsis consists of three components. Merchant introduces an elliptical feature (henceforth, E) to the syntactic derivation and proposes the E feature enters the derivation with a functional head, either C or Voice. At the LF interface, the E feature is interpreted such that the proposition in the ellipsis site mutually entails the one in the antecedent, modulo ∃-type shifting, which basically constitutes e-GIVENness conditions between them, as in (3). (3)

e-GIVEN An expression E counts as e-given iff E has a salient antecedent A and, modulo ∃-type shifting, (i) A entails F-clo (E), and (ii) E entails F-clo (A) (Merchant 2001: 26)

Ha The E feature at the PF interface is interpreted as an instruction that the linearized material c-commanded by the E feature must be unpronounced. Let us consider how ellipsis is licensed in detail with the example in (2). When the E feature enters the derivation in C, as shown in (4), the embedded TP in the second conjunct is taken as an elliptical phrase. The e-GIVENness condition imposes mutual entailment relationship between the TPs via ∃-type shifting. The indefinite NP undergoes existential-closure (4a), and so does the wh-trace in the elided TP (4b). Since there is no focused constituent in the TPs, F-closure of (A) and (E) also gives the same semantic formula in (4c). Now that TPA entails F-clo (E) and TPE entails F-clo (A), it satisfies the e-GIVENness condition in (3). Therefore, the TP she has to pick t in the second conjunct is e-GIVEN and licensed to be elided. (4)

Amanda has to pick one contestant, but she hasn’t decided who [C[E] ]. a. TPA = ∃x [Amanda has to pick x]. b. TPE = ∃x [Amanda has to pick x]. c. F-clo (A) = F-clo (E) = ∃x [Amanda has to pick x]. Now here are puzzles. Not all phrases satisfying the licensing conditions – i.e. e– could be elided. Let us first observe the following examples in (5-6).

GIVENness

(5)

They said they heard about a Balkan language, but I don’t know a. which they said they heard about. No ellipsis b. which. Sluicing c. *which they did. VP-ellipsis

(6)

They attended a lecture on a Balkan language, but I don’t know a. which they attended a lecture about. No ellipsis b. which. Sluicing c. *which they did. VP-ellipsis (Merchant 2008b: 139)

A quick observation regarding these examples tells us that sluicing is possible, but VPellipsis is impossible. This is surprising, given that both VP-ellipsis and sluicing are eligible to be licensed. For instance, suppose that the elided material is VP hear about a Balkan language in (5). Since the elided VP is e-GIVEN by (3), nothing would prevent VP from being elided. The same holds for (6c). A fair amount of literature has attempted to capture why VP-ellipsis is blocked when sluicing is available and proposed the MaxElide constraint (Sag 1976, Williams 1977, Lasnik 2001, Takahashi and Fox 2005, Hardt 2006, Merchant 2001, 2008, among others). (7)

MaxElide (Merchant 2008b) Ellipsis must target the largest constituent which contains an A’-trace.1 1

Merchant (2008b) argues that containing an A’-trace prerequisites for MaxElide to apply. He showed examples below where the smaller constituent can be elided if it does not contain an A’-trace. If

Elliptical Feature Agreement With the constraint in (7), the unavailability of VP-ellipsis in the above examples seems to follow. Wh-movement leaves a trace inside the potential ellipsis site, so the elided target would have an A’-trace in (5-6). This is the environment where MaxElide could be applied. Since the ellipsis target must contain the largest possible constituent and TP is a larger constituent than VP, VP-ellipsis is blocked. Although the MaxElide constraint can account for when VP-ellipsis is unavailable, it does not give us a theoretical reason why this has to be: namely, why such a constraint as MaxElide exists. In the next section of this paper, I will propose a syntactic analysis of this phenomenon with locality of the E feature agreement. There is another puzzling set of examples which Merchant’s E-feature analysis does not seem to capture. Let us compare (8) with (9). (8) is a typical control construction and (9a-b) are ECM and raising structures. (8)

Jane was not sure he could leave, but he tried to <>.

(9)

a. b.

*John believed Mary to know French, but Peter believed Jane to <>. ?? I know that Kate is likely to win the election, but I’m not supposed to reveal that she is likely to <>.

Note that the ECM and raising sentences do not allow VP-ellipsis, but the control construction does (Martin 1996, 2001; Bošković 1997; Koizumi 1995). Assuming that the infinitive T may bear an E feature and license VP-ellipsis, it is unclear what prevents the VP in (9a-b) from being elided. In the following sections, I propose that a functional projection Ellipsis Phrase exist in the left-periphery, and that the unvalued E feature in the E head probe the goal in its c-command domain. Once the feature is valued, it is interpreted at the interfaces. The current analysis provides a syntactic way of accounting for the MaxElide effects, and it also captures why VP-ellipsis is not possible when to-infinitival is selected by a raising or ECM predicate. 2. 2.1

Ellipsis Phrase In the left-periphery

The previous analyses have not clarified the role of syntax for ellipsis phenomena. Merchant (2001), for example, stipulates that an E feature enters the derivation when ellipsis takes place in the sentence. If no ellipsis is involved, no E feature enters the derivation. An immediate question arises as to how the computational system would know whether ellipsis occurs in the sentence, so that the corresponding E feature is included in the numeration. MaxElide operated on (ia-b), (ib) would be blocked; the lower VP is elided even if the higher VP is eligible to be elided, as seen in (ia). Therefore, MaxElide should not be applied to examples where no A’-trace exists in the elided target. (i)

a. b.

The Professor knows that the dissertation committee invited Kyle, but his student doesn’t. The Professor knows that the dissertation committee invited Kyle, but his student doesn’t know that she did.

Ha To answer this question, I assume that a functional projection, dubbed Ellipsis Phrase (EP), exists in the left-periphery (Rizzi 1997), and an occurrence of ellipsis depends on the existence of EP determines. Suppose that, as proposed by Rizzi (1997), CP domain is split into multiple layers. I propose that EP is one of those layers in CP, as shown in (10). (10)

ForceP 3 : EP 3 E : FinP 3

Given the fact that ellipsis is optional, EP is projected only when the speaker intends to execute ellipsis in the sentence. The complementizer head is known to mark a clause-type (Cheng 1991). We extend Cheng’s clause-type hypothesis for ellipsis clause-typing. This makes sense because the C-domain is where the type of clauses is decided. If C is valued with a declarative, interrogative, or exclamative feature, the type of sentence is determined accordingly. In a similar vein, the existence of the E head types ellipsis in the clause. 2.2 Elliptical Feature Agreement Let us elaborate how EP works in more detail. The EP bears an uninterpretable and unvalued uE feature, probing the E feature entered in the functional heads: i.e. Focus and voice.2 The E feature in the functional heads bear the valued E feature, so that it may serve as the goal for probing. Departing from Merchant’s E feature analysis, we do not stipulate that the E feature is just there in the functional head if and only if ellipsis occurs. We assume that the E feature is not a formal feature. It is a convenient name of a certain syntactic feature that plays a role for ellipsis. uQ[ ] in the Focus head counts as E (cf. Merchant 2001, Pesetsky and Torrego 2007). More concretely, uQ[ ] plays a role for E once it is valued by the wh-phrase. uQ[ ] probes and agrees with the wh-phrase. Suppose that the valued uQval in the Focus head counts as E. Following Pesetsky and Torrego (2007), when the probe is valued, it may be able to serve as a goal for the higher probe. Now, the E feature, which is uQval, is probed by uE, and it values uE. Let us call this elliptical feature agreement. The mechanism is shown in (11).

2

To be more explicit, I will assume that the Focus head bears an E feature for sluicing onwards. It seems to me straightforward since wh-movement targets SpecFocusP,

Elliptical Feature Agreement (11)

EP 3 E[uE:Focus] FocusP 3 Focus’ 3 Focus [uQ:2] TP 3 [E] VP 3 WH[2]

uQ[2] is equivalent to the E feature which values uE in the higher position as uE[Focus]. Once the uE feature is valued and the uninterpretable features are deleted, the E feature in the Focus head is licensed to be interpreted at the interfaces; namely, at PF, the TP is instructed to be unpronounced, and at LF, e-GIVENness must be established for this derivation to converge. In other words, ellipsis is possible (i.e. licensed) just for the sister XP of the E feature that enters the elliptical feature agreement relationship with uE. This assumption will play an important role for the current analysis accounting for MaxElide effects which immediately follows this section. Let us consider circumstances where sluicing does not occur. First, suppose that the agreement between wh-phrase and uQ[ ] is established, but no EP is projected. Recall that EP is optional. If this is the case, we assume that the valued uQval does not have to be considered the E feature. Second, if wh-phrase is not numerated or it does not value uQ[ ], the E feature does not exist. Thus, sluicing cannot occur. For example, whether does not value uQ[ ], which in turn means the E feature in the Focus head does not exist, so sluicing is impossible, as seen in (12). (12)

*John claimed he met a Red Sox player last night, but he can’t prove [EP uE[ ] [CP whether .

Let us turn to VP-ellipsis. The E feature in the voice head is equivalent to [voice feature]. Voice has a binary value [active/passive] and the sentence would count as an active or a passive sentence depending on which value the voice head bears. All VPs are assumed to have either of the voice value. I propose that the binary [voice] feature serves as the E feature for VP-ellipsis, as shown in (13) (cf. Merchant 2008a).

Ha (13)

EP 3 E[uE:Voice] TP 3 voiceP 3 voice [active] VP 6 [E]

Suppose that EP is merged. The unvalued uE feature probes the [voice] feature, which is E[Voice]. The E[Voice] feature values uE. Agreed with uE, E[Voice] are interpreted at the interfaces, so that the sister of the voice head would be elided. Again, if EP is not merged, the [voice] feature is not turned into [E]. Thus, no VP-ellipsis occurs. To sum up, ellipsis is optional. In many cases, the speaker has a choice in applying ellipsis in the sentence. The C-domain is the place where a clause is typed, and this is the place where the speaker’s intention is reflected. We assume that the speaker’s intention of using ellipsis must be marked as “ellipsis clause”, so we have created a new functional projection EP bearing an unvalued uE feature. Ellipsis is licensed only if the EP exists and agrees with the E feature in the functional head. The valued E feature is either the valued Q feature in Focus or the [voice] feature in Voice. When the elliptical feature agreement is established, the E feature is interpreted at the interfaces, which means the sister XP of the feature would be unpronounced at PF. We have founded the theoretical backgrounds on ellipsis feature agreement. In the next section, we will explorer how the current analysis accounts for the puzzles introduced earlier: MaxElide and asymmetry of VP ellipsis in to-infinitives. 3. 3.1

MaxElide Deriving MaxElide effects

Now it is time to discuss why VP-ellipsis is blocked when sluicing is possible – the phenomenon described with the constraint called “MaxElide” – as we observed in (5-6), repeated here in (14-15). (14)

They said they heard about a Balkan language, but I don’t know a. which. Sluicing b. *which they did. VP-ellipsis

(15)

They attended a lecture on a Balkan language, but I don’t know a. which. Sluicing b. *which they did. VP-ellipsis (Merchant 2008b: 139)

Under the elliptical feature agreement analysis, the MaxElide phenomenon is straightforward. As commonly assumed in the minimalist framework, we assume that the

Elliptical Feature Agreement only closest E feature agrees with uE (Chomsky 2000). The locality condition on feature agreement captures the empirical fact that VP-ellipsis is blocked whenever sluicing is possible. This is so because uE always agree with the E feature in the Focus head, as shown in (16). (16)

Locality of the E feature agreement EP 3 uE[:Focus] FocusP 3 Focus[E] 3 VoiceP 3 Voice[E] VP

Let us return to our examples in (14-15) and see if this new analysis can account for why sluicing must be applied. In both examples, the TP and VP are e-GIVEN, so that the E feature enters both Focus and Voice. However, the closest head to uE is Focus[E], so [E:Focus] values uE which turns to uE[:Focus]. The E feature that establishes the E feature agreement with uE has an PF-deletion effect: only Focus[E] deletes the TP. Voice[E] fails to value uE, hence no PF-deletion of the VP, explaining why sluicing is derived and VP-ellipsis is blocked. (17) shows how TP-deletion, but not VP-deletion, is derived in (15). (17)

EP 3 uE[:Focus] FocusP 3 DP Foc’ which 3 Focus[E] 6 they attend a lecture on t

The locality of the elliptical feature agreement between uE and the E feature derives MaxElide effects observed in (5-6). The syntactic analysis of MaxElide effects is summarized in (18). (18)

Derivation of MaxElide effects XP is elided at PF if i) there is an EP in the same clause that dominates XP, ii) XP is e-GIVEN, and iii) the sister head of XP bears an E feature which agrees with uE.

Ha 3.2

Optional EPs

Under the assumption that the presence of EPs is optional, we can account for various possibilities of VP-ellipsis in (19). The matrix VP is elided in (19a), while the embedded VP is elided in (19b). (19)

a. b. c.

Ben knows that she invited Klaus, but her father doesn’t . Ben knows that she invited Klaus, but her father doesn’t know she did . Ben knows that she invited Klaus, but her father doesn’t know she invited Klaus.

Let us first consider (19a) where the matrix VP is elided. Suppose that there is an EP in the matrix clause. The matrix VP is e-GIVEN, so the matrix Voice[E] agrees with uE. The matrix VP-ellipsis is derived as in (20). (20)

Ben knows that she invited Klaus, but [EP uE [TP her father doesn’t Voice[E] ]].

I argue that (19b) is derived when there is an EP in the embedded clause, but not in the matrix clause. Recall that EP is projected in the left-periphery, so it may be able to be projected in every clause. The embedded VP is e-GIVEN, and the embedded Voice[E] agrees with uE in the embedded EP, as shown in (21). (21)

Ben knows that she invited Klaus, but [CP1 ø [TP1 her father doesn’t know [EP uE [TP2 she did Voice[E] ]]]].

(19c) is derived when there is no EP in the sentence. This is shown in (22). Every constituent, therefore, must be pronounced. (22)

Ben knows that she invited Klaus, but [CP1 ø [TP1 her father doesn’t know [CP2 ø [TP2 she invited Klaus]]]]. Let us reconsider (5), repeated here in (23).

(23)

They said they heard about a Balkan language, but I don’t know a. which they said they heard about t. b. *which they did . VP-ellipsis c. which . Sluicing

Sluicing over VP-ellipsis in the matrix clause is straightforward. (23c) is derived when the matrix clause has an EP. The matrix Focus[E] agrees with uE, which results in sluicing, and the matrix VP-ellipsis is blocked in (23b), since Voice[E] cannot agree with uE. Recall that I have just argued EP is optional in the clause. Then, we also predict

Elliptical Feature Agreement examples like (24) to be possible, contrary to fact. (24)

They said they heard about a Balkan language, but I don’t know a. *which they said . b. *which they said they did

Let us consider what is predicted to happen in (24a-b). The wh-phrase stops by the specifier position in the intermediate CP on its way of movement to the final destination. Since ellipsis occurs in the lowest clause in both (24a) and (24b), only the embedded clause has an EP (i.e. no ellipsis in the matrix clause). uE in the embedded clause has to agree with Focus[E]. Therefore, deletion of the embedded TP is predicted to be possible in (25), contrary to fact. (25)

*[ø [CP which they said [EP uE[:Focus] [FocusP t’ Focus[E] ]]]].

Although it is unclear why, it seems to be correct to assume that the SpecCP position must be overtly pronounced to license sluicing in English. The embedded TP with a covert wh-phrase in the clause would never undergo sluicing. Let us construct a hypothesis of pronunciation of wh-phrase in sluicing in (26). (26)

Overtness of wh-phrase in sluicing Sluicing is licensed in CP whose specifier is overtly realized with wh-phrase.

The constraint in (26) rules out (25), where the intermediate SpecFocusP is occupied by a covert wh-operator. That is why (24a) is not possible. It is important to note that the Focus head still values uE. Although Focus[E] would license sluicing, ellipsis is unavailable for an independent reason (26). This prevents the next closest head Voice from agreeing with uE for VP-ellipsis in (26). And this is why (24b) is not possible. (26)

*[ø [CP which they said [EP uE [CP t’ they did Voice[E] ]]]].

4.

Asymmetry of VP-ellipsis in to-infinitives

In section 1, we introduced an asymmetry of VP-ellipsis in to-infinitival constructions. We observed that VP-ellipsis is possible when to-infinitive is selected by a control predicate, but VP cannot be elided if to-infinitive is selected by either a raising or an ECM predicate. Examples are repeated here in (27-28). The previous analyses within the minimalist framework would have to stipulate that the E feature can enter the derivation with a control to-infinitive, but it cannot with raising or ECM to-infinitives. (27)

Jane was not sure he could leave, but he tried to <>.

(28)

a. b.

*John believed Mary to know French, but Peter believed Jane to <>. ?? I know that Kate is likely to win the election, but I’m not supposed to reveal that she is likely to <>.

Ha Under the elliptical feature agreement analysis, the difference follows straightforwardly. The core difference between the two types of to-infinitives is whether CP is involved. Since EP is merged within CP, the existence of CP in the to-infinitive structure is crucial for VP-ellipsis. In control, this is possible, but in raising or ECM, it is impossible. Let us analyze the examples again. In (29), the higher verb try selects a control predicate, so the embedded clause must be a CP. For us, this means EP can be projected in the embedded clause. Suppose that EP exists. The E feature in the Voice head values uE, hence licensing VP-ellipsis. (29)

Jane was not sure he could leave, but he tried [EP uE[:voice][CP [TP PRO to [voiceP voice[E] [vP ]]]]].

In (30a-b), TP is the highest projection in the embedded clause since the higher verbs believe and adjective likely select an ECM and a raising construction respectively. In (30a), the lowest EP that may be projected is the matrix clause in the second conjunct. uE probes the closest [voice] feature. Note that there are two Voice heads in its ccommand domain, and the one that agrees with uE would be the higher [voice] feature. Since the lower [voice] feature fails to agree with uE, the lower VP-ellipsis is not licensed. The same holds for (30b). The that-clause is the first place where EP could be projected. Inside the that-clause, there are two Voice heads. By locality of the elliptical feature agreement, the higher [voice] feature values uE. Since the lower [voice] feature does not agree with uE, VP-ellipsis, shown in (30b) is not licensed. (30)

a. b.

*John believed MARY to know French, but [EP uE[:voice1] [TP Peter [VoiceP1Voice1[E] believed JANE [TP to [VoiceP2 Voice2[E] .]]]]] ?? I know that Kate is likely to win the election, but I’m not allowed to reveal [EP uE[:voice1] [CP that [TP she is [VoiceP1 Voice1[E] likely [TP to [VoiceP2 Voice2[E] .]]]]]]

If the higher [voice] feature agrees with uE, it should be possible to elide the higher VP. This prediction is borne out in (31). (31)

I know that Kate is likely to win the election, but I’m not allowed to reveal that she is <>.

5.

Conclusion

In this paper, we have been concerned with two puzzles on ellipsis: MaxElide and asymmetry of VP-ellipsis in to-infinitive constructions. The current analysis, using elliptical feature agreement, provides theoretical solutions for them. We assume that Ellipsis Phrase is projected in the left-periphery and it bears an unvalued uE feature. uE searches for the goal to be valued and a certain functional head, such as Focus or Voice may bear the valued E feature. We applied locality condition to the elliptical feature agreement. If there is more than one E feature downstairs which uE is able to agree with, the closest E feature values uE. That is how we capture MaxElide effects. Since the E

Elliptical Feature Agreement feature licensing sluicing is the Focus head, and it is closer to uE than the one licensing VP-ellipsis which is the [voice] feature in the Voice head, VP-ellipsis is predicted to be blocked whenever sluicing is available. The asymmetry in to-infinitive construction is also straightforward. The internal structure inside the to-infinitival clauses differs; the one type involves control, and the other involves ECM or raising. The crucial difference between them would be whether the to-infinitival clause is a CP or a TP. This is important because EP can only be projected within CP. This captures the fact that VPellipsis is possible in the control to-infinitive structure, but impossible in the raising toinfinitive structure.

References Bošković, Ž. 1997. The Syntax of Nonfinite Complementation. Cambridge, MA.: MIT Press. Cheng, L. 1991. On the Typology of Wh-questions. PhD dissertation, MIT. Chomsky, N. 2001. Minimalist Inquiries: The Framework. In R. Martin, D. Michaels, and J. Uriagereka (eds.), Step by Step: Essays on Minimalist Syntax in Honor of Howard Lasnik. Cambridge, MA: MIT Press, 89–155. Hardt, D. 2006. Re-binding and the Derivation of Parallelism Domains. Proceedings of BLS 32. Koizumi, M. 1995. Phrase Structure in Minimalist Syntax. Ph.D. dissertation, MIT. Lasnik, H. 2001. When Can You Save a Structure by Destroying it? In M., Kim, and U., Strauss (eds.), Proceedings of NELS 31, Georgetown University: GLSA, 301-320. Lobeck, A. 1995. Ellipsis: Functional Heads, Licensing and Identification. New York: Oxford University Press. Martin, R. 1996. A Minimalist Theory of PRO and Control. Doctoral dissertation, University of Connecticut, Storrs. Martin, R. 2001. Null Case and the Distribution of PRO. Linguistic Inquiry 32: 141-166. Merchant, J. 2001. Syntax of Silence: Sluicing, Islands, and the Theory of ellipsis. Oxford:Oxford University Press. Merchant, J. 2008a. An Asymmetry in Voice Mismatches in VP-ellipsis and Pseudogapping. Linguistic Inquiry 39: 169-179. Merchant, J. 2008b. Variable Island Repair Under Ellipsis. In K. Johnson (ed.), Topics in Ellipsis. Cambridge: Cambridge University Press, 132-153. Pesetsky, D., and E. Torrego. 2007. The Syntax of Valuation and the Interpretability of Features. In S. Karimi, V. Samiian, and W. K. Wilkins (eds.), Phrasal and Clausal Architecture. Amsterdam: John Benjamins, 262-294. Rizzi, L. 1997. Locality and Left-periphery. In L. Haegeman (ed.), Elements of Grammar, Dordrecht: Kluwer, 281-337. Sag, I. 1976. Deletion and Logical Form. Ph. D. Dissertation, MIT. Takahashi, S., and D. Fox. 2005. MaxElide and the Re-binding Problem. Proceedings of SALT 15. Williams, E. 1977. Discourse and Logical Form. Linguistic Inquiry 8: 101-139.

Ha Department of Linguistics Korea University Seoul 136-701 Korea [email protected]