![Sample our Bioscience journals, sign in here to start your access, Latest two full volumes FREE to you for 14 days]({"type":"image" , "src" : "/sda/5925/TOC-Subject-Sample-2021-Bioscience.jpg"})
ABSTRACT
Implicit expectations play a central role in sentence processing. These expectations are often assumed to be static or change only at relatively slow time scales. Some theoretical proposals, however, hold that comprehenders continuously adapt their expectations based on recent input. Existing evidence has relied heavily on self-paced reading, which requires familiarisation with a novel task. We instead employ eye-tracking reading to investigate the role of expectation adaptation during speeds and task demands more closely resembling natural reading. In two experiments, subjects read sentences that contained higher than expected proportions of a previously highly unexpected structure (reduced relative clauses). We test how this change in the statistics of structures within the experiment affects reading: if subjects adapt their expectations, reading times for the unexpected structure should decrease over the course of the experiment. This prediction is confirmed in both experiments. Significant effects of the changing statistics are observed for regression-related measures but not first-pass reading measures. We discuss possible accounts of this pattern in the eye-movement record.
Acknowledgements
Parts of this work were presented at the 36th Annual Conference of the Cognitive Science Society (Farmer et al., Citation2014). The authors thank members of the Human Processing Lab for providing feedback on earlier presentations of this work. We are particularly indebted to Thomas Farmer who conceived this project, guided design decisions, supervised data collection, and was central in shaping the theoretical perspectives discussed here. The original manuscript contained a mistake in Study 3 that we would not have become aware of without the thorough work of an anonymous reviewer. The studies presented here were partially supported through NICHD R01 HD075797. The views presented here do not necessarily reflect those of the funding agencies. All stimuli, data, and analysis scripts are available via OSF https://osf.io/5bpzh/.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Notes
1 Not all accounts of garden path effects reference expectations or computationally related concepts, such as gradient constraints or contextually-cued memory retrieval (e.g., Levy, Citation2008; Lewis et al., Citation2006; for review, see Gibson & Pearlmutter, Citation1998). One alternative are re-analysis and other two-stage accounts (e.g., Frazier & Fodor, Citation1978), which we return to in the general discussion. To the extent that garden path effects change throughout the course of the experiment, and do so differently for MVs and RCs, this requires explanation in any account of garden paths (expectation adaptation constitutes one such explanation). The question we raise here is thus of relevance to expectation-based theories of sentence processing.
2 Fine et al. (Citation2013) reported relative probabilities of .008 for RCs and .7 for MVs based on verb subcategorization database from Roland et al. (Citation2007). These estimates were obtained by only conditioning the ambiguous verb form – i.e., p(structure | “warned”) for the example in (1). This ignores other information available to the reader prior to the disambiguation region (e.g., the prepositional phrase about the dangers in (1)). We thank an anonymous reviewer for making us aware of this limitation. The estimates used here are based on the same corpus (the British National Corpus, British National Corpus Consortium, Citation2007) that was also employed by Roland et al. (Citation2007). Unlike Fine et al. (Citation2013), however, we take into account the additional material prior to the point of disambiguation (for details, see Appendix A). The critical prediction tested here – the significant reduction of the ambiguity effect on RCs with increasing exposure to RCs – holds under a broad range of plausible ways of estimating the relative probability of RCs in everyday language experience: in everyday language use, RCs are orders of magnitude less likely than MVs in the type of garden path environment shown in (1).
3 Complete adaptation is not predicted under the hypothesis of expectation adaptation (Fine et al., Citation2013): ideal adaptation requires comprehenders to trade off flexibility and stability in their beliefs about the statistics of the input (Kleinschmidt and Jaeger, Citation2015, Part II). Bayesian belief-updating provides one model of incremental expectation adaptation, and has been shown to predict human reading data (see Fine et al., Citation2010; Jaeger et al., Citation2019; for a related model, see also Chang et al., Citation2006). Critically, adapted expectations will always change more quickly for the less expected structure, in particular under the surprisal link (for demonstration, see Jaeger et al., Citation2019). The qualitative predictions described here for the hypothetical case of complete adaptation thus generalise to the more realistic case of partial adaptation.
4 Previous research has sometimes focused on first-pass measures in order to show that expectations can affect the earliest moments of sentence processing (e.g., Garnsey et al., Citation1997; Smith & Levy, Citation2013; Trueswell et al., Citation1993). This differs from the focus of the present work: expectation-based processing does not necessarily result in anticipatory behaviour (for discussion, see Kuperberg & Jaeger, Citation2016; Levy, Citation2008; Tanenhaus, Citation2004).
5 We thank an anonymous reviewer for alerting us to this possibility. This unintentional property of fillers applies to both Experiment 1 and 2.
6 We note that computational models of expectation adaptation predict superlinear effects of item order on the ambiguity effect, in particular for RCs (Fine et al., Citation2010; Jaeger et al., Citation2019). As computational modelling is beyond the scope for the present work, we facilitate comparison to Fine et al. (Citation2013), and test linear effects. This common simplifying assumption is not expected to bias in favour of the hypothesis of expectation adaptation.
7 We re-specifying the regression equation as DV ∼ Structure * Ambiguity * Item Order – Ambiguity + log(stimulus order) + Number of characters + Random Effects. This yields the simple effects of ambiguity for both MVs and RCs. All of our simple effect analyses follow the same approach.
8 R formula: RT ∼ Paradigm + s(TrialOrder) + s(TrialOrder, by = Paradigm) + s(Subject, bs=‘fs’) + s(Item, bs=‘fs’).
9 It is possible that other distributional assumptions and link functions would be more adequate for reading data, and that our results might not hold under these alternative approaches. This is an active area of research (e.g., Rouder, Citation2019; Wagenmakers & Brown, Citation2007). Here we have limited ourselves to the two most commonly employed link functions for reading times (normally and lognormally distributed RTs).
10 Similarly, only one of the analyses of Experiment 2 found significant trial effects (second-pass reading times). We focus here on Experiment 1 because Experiment 2 differed from previous self-paced reading experiments in that the critical lexical material was repeated across stimuli.
11 The magnitude of k-way interactions does not necessarily constrain the statistical power to detect higher-level interactions containing the k-way interaction. However, the three-way interaction predicted by the hypothesis of expectation adaptation is one in which the ambiguity effect for RCs is reduced towards 0 (rather than inverted). That is, if the hypothesis of expectation adaptation has merit, we expect the magnitude of the ambiguity effect for RCs (as reflected in the two-way interaction of ambiguity and structure) to constitute a bound on the power to detect a significant three-way interaction of ambiguity, structure, and item order. This possibility is what we are discussing here.