Passives are not hard to interpret but hard to remember: evidence from online and offline studies

ABSTRACT

Passive sentences are considered more difficult to comprehend than active sentences. Previous online-only studies cast doubt on this generalisation. The current paper directly compares online and offline processing of passivization and manipulates verb type: state vs. event. Stative passives are temporarily ambiguous (adjectival vs. verbal), eventive passives are not (always verbal). Across 4 experiments (self-paced reading with comprehension questions), passives were consistently read faster than actives. This contradicts the claim that passives are difficult to parse and/or interpret, as argued by main perspectives of passive processing (heuristic, syntactic, frequentist). The reading time facilitation is compatible with broader expectation/surprisal theories. When comprehension targeted theta-role assignment, passives were more errorful, regardless of verb type. Verbal WM measures correlated with the difference in accuracy, but not online measures. The accuracy effect is argued to reflect a post-interpretive difficulty associated with maintaining/manipulating the passive representation as required by specific tasks.

KEYWORDS:

• language comprehension
• passivization
• heuristics
• surprisal
• self-paced reading

Acknowledgments

This research was partly funded by the DFG – Leibniz Prize AL 554/8-1 awarded to Artemis Alexiadou. We gratefully acknowledge the DFG contribution. We would also like to gratefully thank John Hale for assistance with the Brown Corpus search.

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Caterina Laura Paolazzi http://orcid.org/0000-0002-9975-3883

Notes

1. We only consider the plausible passives (either arguments could be agent or patient) given our interest is the complexity of passive syntax/word order. However, a similar difference was found with the implausible conditions.

2. Model did not converge with random slopes. The same model was used for the analysis of theta-role questions only and for first- vs. second-half analysis, given that more complex models did not converge.

3. Model only converged with random slope for subjects but not for items. The same model was used for the analysis of theta-role questions only and for first- vs. second-half analysis, given that more complex models did not converge.

4. Data from trials 7–51 (the first 6 items were practice items) were analysed as “first-half” and data from trials 51–96 as “second-half”.

5. The data were divided in first vs. second half as per Experiment 1.

6. For all the reading times analyses, the most complex model, including both intercept and random slope for both subjects and items, always converged. The only exception was the first adjective region, where the model only converged with intercept for both subjects and items and slope only for subjects and not items.

7. Model only converged with intercepts.

8. Data from trials 7–51 (the first 6 items were practice items) were analysed as “first-half” and data from trials 51–96 as “second-half”.

9. Model for the verb region analysis only included Syntax in the structure of the random effect of the items, as the verb differed across predicate type.

10. For all accuracy analyses, the model only converged with random intercepts and not slopes.

11. The model only converged with syntax and not predicate type in the structure of the Item random effect.

12. The data were divided in first vs. second half as per Experiment 1, 2 and 3.

13. The presence of a silent argument in short passives is supported by several syntactic diagnostics, including the ability to support subject controlled infinitival sentences and subject-oriented modifiers and depictives (e.g. “The book was written to collect the money/deliberately/drunk”), and to bind reflexives (e.g. “such privileges should be kept to oneself”).

14. This perspective seems as though it could also be accounted for under a surprisal based account.

Additional information

Funding

This research was partly funded by the DFG – Leibniz Prize AL 554/8-1 awarded to Artemis Alexiadou. We gratefully acknowledge the DFG contribution. We would also like to gratefully thank John Hale for assistance with the Brown Corpus search.