![Sample our Language & Literature journals, sign in here to start your FREE access for 14 days]({"type":"image" , "src" : "/sda/5941/TOC-Subject-Sample-2021-Language-Literature.jpg"})
Abstract
Interpretation of a pronoun in one clause can be systematically affected by the verb in the previous clause. Compare Archibald angered Bartholomew because he … (he = Archibald) with Archibald criticised Bartholomew because he … (he = Bartholomew). While it is clear that meaning plays a critical role, it is unclear whether that meaning is directly encoded in the verb or, alternatively, inferred from world knowledge. We report evidence favouring the former account. We elicited pronoun biases for 502 verbs from seven Levin verb classes in two discourse contexts (implicit causality and implicit consequentiality), showing that in both contexts, verb class reliably predicts pronoun bias. These results confirm and extend recent findings about implicit causality and represent the first such study for implicit consequentiality. We discuss these findings in the context of recent work in semantics, and also develop a new, probabilistic generative account of pronoun interpretation.
Acknowledgements
Thanks are due to Jesse Snedeker, Jess Sullivan, Hugh Rabagliati, Pierina Cheung, Jennifer Arnold, Jennifer Rodd, Mahesh Srinivasan and Andrew Stewart for comments and suggestions, to John H. Krantz for assistance in subject recruitment, and to the volunteers who participated in this study.
Disclosure statement
No potential conflict of interest was reported by the authors.
Notes
1. Note that ‘implicit causality’ is also used to refer to a task in which people make inferences about various attributes of individuals based on events they have participated in (Brown & Fish, Citation1983a). Though initially thought to be related, subsequent research has found little or no relationship between the two kinds of implicit causality (Hartshorne, Citation2014). All discussion of implicit causality in the present paper pertains to the original Garvey and Caramazza phenomenon.
2. An alternative is that these biases are in fact learned heuristics derived from the statistics of pronoun use itself, which are used to predict the likely reference of a pronoun (Crawley et al., Citation1990; Fletcher, Citation1984). We return to this account in the General Discussion.
3. As indicated in the citations, Brown and Fish (1Citation983b) have been influential in the development of both positions and can be read as supporting either one.
4. We thank Andrew Stewart for pointing this out.
5. For instance, the reason that both Agnes broke the vase and The vase broke are grammatical but Beatrice hit the vase is grammatical while *The vase hit is not is that break describes an externally caused event, whereas hit does not (cf. Levin & Rappaport Hovav, Citation2005).
6. In the psycholinguistic literature, emotion verbs have often been grouped together with propositional attitude verbs (think, believe) and education verbs (teach, learn). Collectively, these verbs are known as ‘psych verbs’. Because these verbs appear in distinct verb classes, we discuss them separately. Moreover, H&S found little evidence that these different types of psych verbs pattern similarly with regard to re-mention biases.
7. We thank Jennifer Arnold for raising this point.
8. A scale is linear if a change of d units has equal value at any point in the scale. Percentages violate this rule: The difference between 50% and 51% (d = 1%) is less meaningful than the difference between 99% and 100% (d = 1%). This complicates comparison of results, since large numerical differences can actually be ‘smaller’ than small numerical differences. For demonstration and discussion, see Jaeger (Citation2008).
9. If not correcting for multiple comparisons, classes 45.4 and 59 are also significantly different from one another (t(134) = 2.5, p = .01).
10. We assessed significance with a permutation analysis: We randomly reassigned verbs to verb class with the constraint that the number of verbs in each class remain the same, and then refit the model. We repeated this process 1000 times. In all 1000 iterations, the standard deviation of the random intercepts and slopes was much greater than the true model, never dropping below 1.15 and 1.58, respectively.
11. Causes of an event (Archibald angered Bartholomew) include both the power of the agent to realise some effect (causing anger) and the liability of the patient to be affected (the ability to be angry; cf. White, Citation1989). Moreover, there are proximate causes (Archibald punching Bartholomew in the nose) and prior causes (Archibald’s troubled upbringing, which made him violent).
12. The study was conducted on Amazon Mechanical Turk and participants received monetary compensation. An additional five participants were excluded for answering all four unambiguous filler trials incorrectly, nine were excluded for failing to answer every question, and three were excluded for not being native English speakers. Whether the subject was male and the object was female or vice versa was counterbalanced within and between participants. Two orders of stimuli were used, one of which was the reverse of the other. Half the participants saw the verbs in present tense, half in past tense.
13. Participants and verbs were random effects, while tense and whether the male character was the subject or object were fixed effects (the latter was also non-significant, Wald’sz = 1.65, p = .10). Maximal random effects structure was used.
14. Here we use the term generative model in the original sense of a model that explains how some observable behaviour was generated, and not in reference to a particular tradition of syntactic theory (cf. Tenenbaum, Kemp, Griffiths, & Goodman, Citation2011).
15. We are indebted to Sagi and Rips (Citationin press) for the suggestion of using the probabilities of the events themselves. Among the many differences between our approach and theirs is that rather than invoking Gricean reasoning, Sagi and Rips embed their theory in the notion of causal identity: What makes he in likely to refer to Archibald in (6) is the fact that he and Archibald are likely to be causally related. We are currently attempting to tease apart these two accounts in ongoing research.
16. Because most classes do not lend themselves to intuitive names, these classes are numbered. Throughout, we use the VerbNet numbers and give examples.