Processing VP-ellipsis and VP-anaphora with structurally parallel and nonparallel antecedents: An eye-tracking study

Abstract

In this paper, we report on an eye-tracking study investigating the processing of English VP-ellipsis (John took the rubbish out. Fred did [] too) (VPE) and VP-anaphora (John took the rubbish out. Fred did it too) (VPA) constructions, with syntactically parallel versus nonparallel antecedent clauses (e.g., The rubbish was taken out by John. Fred did [] too/Fred did it too). The results show first that VPE involves greater processing costs than VPA overall. Second, although the structural nonparallelism of the antecedent clause elicited a processing cost for both anaphor types, there was a difference in the timing and the strength of this parallelism effect: it was earlier and more fleeting for VPA, as evidenced by regression path times, whereas the effect occurred later with VPE completions, showing up in second and total fixation times measures, and continuing on into the reading of the adjacent text. Taking the observed differences between the processing of the two anaphor types together with other research findings in the literature, we argue that our data support the idea that in the case of VPE, the VP from the antecedent clause necessitates more computation at the elision site before it is linked to its antecedent than is the case for VPA.

Keywords:

• Sentence processing
• VP-ellipsis
• Eye tracking during reading

Notes

¹In recent theoretical work, for example, Merchant (2007, 2008), Tanaka (2011), it has been argued that one of the “nonparallel” constructions investigated in this paper, namely, the passive (voice mismatch) condition illustrated in (1), is in fact underlyingly syntactically parallel, and that apparent parallelism effects found in traditional judgment data are best accounted for in terms of discourse factors (Kehler, 2000, 2002). While acknowledging the relevance of such work in the overall treatment of VP-ellipsis, it does not, we believe, invalidate any of the findings of the current paper. There are three reasons for this. First, it is unclear to us that the intuitive judgments on which Merchant, Tanaka and others rely have any necessary primacy over the experimental data presented here, which demonstrate a clear behavioural interaction between (surface) parallelism and anaphor type, irrespective of any putative underlying grammaticality: that is to say, surface mismatch implicitly “matters” to readers. Second, we have demonstrated in previous work, using different measures, that antecedents containing voice mismatches are consistently judged less acceptable than surface parallel antecedents with VPE, notwithstanding the fact that this parallelism effect is weaker than with nominal antecedents (Duffield, Matsuo, & Roberts, 2009). Finally, the goals of this paper are primarily psycholinguistic, rather than theoretical, in nature: we are concerned with how readers process different surface strings in real time given varying contexts, rather than with how far behavioural judgments comport with theoretical constructs. Of course, the two concerns should be ultimately related, but such considerations are beyond the scope of this paper.

³Merchant (2007, 2008) discusses the fact that voice-mismatched pseudogapping (e.g., *Roses were brought by some, and others did lilies) is worse than voice-mismatching VPE (e.g., This problem was to have been looked into, but nobody did). The difference is assumed to lie in the size of the elided clause, with VPE targeting sister of Voice, and thus low enough in the syntactic tree to be neutral between active and passive voice, in contrast to pseudogapping (also sluicing and fragment answers), which involves the elision of a constituent higher up in the tree (TP).

²Those papers built on earlier discussion by Postal, Bresnan, Ross, and others, concerning the question of whether ellipsis and (overt) pronominal anaphora should be handled by separate mechanisms—and related to this—the issue of whether overt pronouns were the surface realization of underlying full noun-phrases. Though the latter question appears to have been laid to rest by the late 1970s, the former question is still a matter of some controversy.

⁴For VPE at least, these results are also supportive of recent semantic accounts that assume that VP-ellipsis with voice mismatches are underlyingly grammatical (Merchant, 2007, 2008; Tanaka, 2011; see notes 1 and 11).

⁵Murphy (1985) notes that in VPA, the pronoun refers to a discrete action, entity or event (ib) and is much less ambiguous than VPE (ic), for which there is a larger number of possibilities overall of potential antecedents.

(i) a. Jean saw the accident.

  I thought Chris did.

  *I thought Chris did it.

  b. Does Mary eat meat?

  No, but Ed does.

  *No, but Ed does it.

  c. Jane saw Fred pull the plug.

  I thought Amy did. (ambiguous)

  I thought Amy did it. (unambiguous)

Similarly, in Lasnik's analysis (1999a, 1999b), a distinction is made between main verbs and auxiliaries, with the latter subject to strict formal identity conditions (e.g., *John has left but Mary shouldn't), in contrast to the former, where “sloppy identity” appears to be at work (e.g., John slept and Mary will too). It is not clear from current research how much effect such differences may have on the real-time linking processes in VPA versus VPE.

⁶We thank an anonymous reviewer for this suggestion.

⁷Although the critical segment was indeed one word longer in the VPA conditions, the pronominal was fixated only 25% of the time (as is often the case when reading “functional” versus lexical elements, Rayner, Juhasz, & Pollatsek, 2005). When these fixations are not included, the results remain the same.

⁸We thank an anonymous reviewer for this suggestion.

⁹This analysis of VPE is different to that of Merchant (2007, 2008), in that Arregui et al. assume that VPE with a voice-mismatched antecedent is ungrammatical (and the parser works in grammatically constrained steps to ensure a syntactically parallel structure) whereas in Merchant's account, such voice-mismatches should be grammatical (see also Tanaka, 2011; and note 1 above). Taking these processing findings and those that show that VPE with nonparallel voice-(mismatch) antecedents elicit degraded acceptability judgments (e.g., Duffield, and colleagues; Kim & Runner, 2009) there is clearly a processing cost at the elision site. What exactly occurs is open to interpretation. If the Tanaka/Merchant account is correct, this would involve a change of feature-values in the [voi] head, and not reconstruction, which would be the case if one assumes the traditional, underlyingly ungrammatical analysis for nonparallel VPE. A future study focusing on the graded acceptability of VPE versus pseudogapping constructions would prove highly interesting.

¹⁰One reason that our results cannot arbitrate between these two accounts of VPE processing is that our materials crucially differ from those of Martin and McElree. Specifically, the processing differences discussed above relate to the processing of VP-ellipsis with syntactically nonparallel antecedents (passive vs. nominal), unlike those in the Martin and McElree study. Before interpretation can take place, the comprehender must somehow fix the ungrammaticality and this appears to take more time in the simpler passive nonparallel versus the nominal nonparallel conditions.

¹¹As pointed out by an anonymous reviewer, another important influencing factor on VPE, but one that we did not examine in the current study, is different discourse relations between the antecedent and the elided clause, following, for instance, Kehler's (2000, 2002) Coherence Theory (see also Simner, Pickering, & Garnham, 2003, for do it anaphors). Kehler's is a hybrid approach to VPE, where under some discourse conditions, syntactic identity is required (Resemblance relations, e.g., *Jill betrayed Abbey, and Matt was too), but for others, the antecedent and elided clause need only to match at the propositional level (Cause-Effect, e.g., The report was critical of Roy, so Kate didn't). In the current study, we kept the discourse relations constant across parallelism and anaphor type conditions, however in terms of the predictions that can be made on the basis of Kehler's account, mixed results have been observed. A study by Frazier and Clifton (2006) specifically tested Kehler's claim that syntactic parallelism is required for Resemblance but not for Cause–Effect discourse relations and found that that parallelism was preferred under experimental conditions for all types of coherence relations. Thus the processing predictions of the theory were not supported. In contrast, Kim and Runner (2009) report that their data from three magnitude estimation studies show that syntactic parallelism effects are indeed modulated by discourse relations. Furthermore, information structure-level factors are also likely to affect VPE and VPA (cf., Kertz, 2008), given that in creating our nonparallel antecedents, a change takes place in the information structure properties of the antecedent clause, as well as its syntactic structure. Clearly more research is needed to investigate the interactions between discourse-level and syntactic factors that affect the processing of both VPE and VPA, and we leave these interesting questions for future research.