Grammatical and information-structural influences on pronoun production

Abstract

A standard assumption in psycholinguistic research on pronoun interpretation is that production and interpretation are guided by the same set of contextual factors. A line of recent research has suggested otherwise, however, arguing instead that pronoun production is insensitive to a class of semantically driven contextual biases that have been shown to influence pronoun interpretation. The work reported in this paper addresses three fundamental questions that have been left unresolved by this research. First, research demonstrating the insensitivity of production to semantic biases has relied on referentially unambiguous settings in which the comprehender's ability to resolve the pronoun is not actually at stake. Experiment 1, a story continuation study, demonstrates that pronoun production is also insensitive to semantic biases in settings in which a pronoun would be referentially ambiguous. Second, previous research has not distinguished between accounts in which production biases are driven by grammatical properties of intended referents (e.g., subject position) or by information-structural factors (specifically, topichood) that are inherently pragmatic in nature. Experiment 2 examines this question with a story continuation study that manipulates the likelihood of potential referents being the topic while keeping grammatical role constant. A significant effect of the manipulation on rate of pronominalisation supports the claim that pronoun production is influenced by the likelihood that the referent is the current topic. Lastly, the predictions of Kehler et al.'s Bayesian analysis of the relationship between production and interpretation have never been quantitatively examined. The results of both experiments are shown to support the analysis over two competing models.

Keywords:

• pronoun production
• implicit causality
• information structure

Acknowledgements

We thank Albert Gatt and three anonymous reviewers for useful comments and criticisms, and Roger Levy for helpful discussion. This research was supported by an Andrew W. Mellon postdoctoral fellowship to the first author. The results of Experiment 2 have been presented at the 22nd Annual CUNY Conference on Human Sentence Processing, the 2011 meeting of the German Linguistics Society, the 2011 Constraints in Discourse meeting and the 86th Annual Meeting of the Linguistics Society of America. We thank our audiences at those meetings for useful discussions, and Brittany Young and Meredith Larson for their help in annotation.

Notes

1. Here and throughout the paper, we use the term “speakers” to include both speakers and writers.

2. And hence, contra a claim found in the psycholinguistics literature (Chambers & Smyth, 1998, inter alia), this algorithm will not always identify the previous subject as the preferred referent.

3. The denominator of (3) is simply the probability that a pronoun is the form of reference chosen by the speaker P(pronoun)), which can be computed by summing the numerator over all referents that are compatible with the pronoun. This term has the effect of normalising the probabilities to 1.

4. Arnold (2001) reported on a story continuation study that included Source-Goal transfer-of-possession contexts like (2). She found the same production asymmetry as Stevenson et al., whereby 76% of references to the subject Source were pronominalised yet only 20% of those to the object-of-PP Goal were. However, the next-mention bias towards the Goal was an overwhelming 86%. Equation (3) would actually predict a 61% interpretation bias to the Goal for her stimuli if a pronoun prompt were to be included (Kehler et al., 2008).

5. Arnold (2001) reported an effect of thematic role on rate of pronominalisation in her study of transfer-of-possession contexts, whereby references to the Goal were pronominalised more often than references to the Source. Fukumura and van Gompel (2010) point out a number of reasons to suggest that Arnold's result is not fully conclusive, however, including the fact that her Source-Goal and Goal-Source contexts differed not only in the order of the thematic role fillers but in a number of other potentially relevant respects as well.

6. This study appeared as Experiment VI in the first author's dissertation (Rohde, 2008). It is presented here with additional analyses pertaining to the Bayesian model.

7. As mentioned earlier, these analyses reflect a conservative data inclusion strategy in which a continuation was excluded if at least one coder assessed it as ambiguous. The pattern of statistical significance for all results remains the same if continuations are included for which at least one coder assigned a non-ambiguous interpretation.

8. As can be seen in Figure 3, the predictions of the Bayesian and Mirror models are very close for the non-IC condition. This reflects the fact that the next-mention biases for this condition were close to 50–50 for the two referents. The two models are equivalent when the prior is uniform.

9. As mentioned above, our account predicts that the correlation between the values produced by the Mirror model and actual interpretation biases are due to the strong effect of grammatical role on production. Hence, we would not expect to find a correlation if we restrict analysis to include only subject or non-subject referents. This prediction is borne out: When the correlation is restricted to one referent only, the R² values drop considerably (by participants: R² = 0.07, F₁(1,61) = 5.377, p < 0.05; by items: R² = 0.09, F₂(1,69) = 8.022, p < 0.01) compared to the values shown in Table 1 in which both referents were included. Further, the remaining significance is likely driven by the fact that not all non-subject referents were direct objects; some were objects of prepositional phrases and hence expected to have an even lower pronominalisation rate than direct objects. In an analysis that excludes those items in which the non-subject was an object-of-PP (almost all the non-IC verbs as well as the IC verbs apologise to, confess to, and stare at), the correlation between the observed data and the Mirror model is no longer significant (by participants: R² = 0.04, F₁(1,40) = 2.865, p = 0.10; by items: R² = 0.01, F₂(1,40) = 1.557, p = 0.22).

10. An alternative to casting the phenomenon in terms of likelihood of being the topic would be to treat topicality itself as a gradient rather than binary concept (Arnold, 2010; Givón, 1983), with grammatical subjects being more topical than objects and so forth. The proposal offered here is compatible with either possibility, and hence we will not attempt to resolve the issue further.

11. One might note that the rate of pronominalisation towards subjects in the active condition (62.1%) appears lower than for the same condition in Experiment 1 (77.5%), which used similar (although not identical) stimuli. Further analysis of the data revealed that seven of the participants in Experiment 2 never used a pronoun in any continuation; this was not the case for any participants in Experiment 1 nor is it typical to see in other experiments. The results after the data for these seven participants is removed are as follows:

Table

The results are now highly consistent with the previous experiment (76.1% v. 77.5% for references to the subject in active subject-biased contexts, and 28.3% v. 26.6% for references to the object in such contexts). The results of statistical analysis with this exclusion match those reported in the main text.

12. Again, the analysis reflects a conservative data inclusion strategy in which a continuation was excluded if at least one coder assessed it as ambiguous. As was the case in Experiment 1, the pattern of statistical significance for all results remains the same if continuations are included for which at least one coder assigned a non-ambiguous interpretation.

13. Caramazza and Gupta (1979) similarly found an effect of passivisation in IC contexts using a timed comprehension task. Kaiser, Cheng-Huan Li, & Holsinger (2011) describe a complementary effect of passivisation in Agent–Patient contexts.

14. An anonymous reviewer remarks that the 59.1% figure in the active condition seems low considering the strong biases usually associated with IC verbs. However, biases previously reported in the literature are almost always collected using prompts containing “because” which, by restricting the continuations to causal follow-ons, enhances the bias towards the causally implicated referent (i.e., the subject for subject-biased IC verbs). Similarly, these prompts also commonly include a subject pronoun, which as we have seen will also raise the subject bias. Indeed, our result replicates the bias reported for a similar condition in Kehler et al. (2008; Experiment 3, page 33, Section 5.1.4).

15. Footnote 9 reported on separate analyses for referents in different grammatical roles that established the lack of a correlation between the predictions of the Mirror model and the observed interpretation biases. Since our analysis predicts an effect of voice on rate of pronominalisation for referents in subject position in this experiment, a similar analysis is not applicable here.