Linguistic Context in Verb Learning: Less is Sometimes More

ABSTRACT

Linguistic contexts provide useful information about verb meanings by narrowing the space of candidate concepts. Intuitively, the more information, the better. For example, “the tall girl is fezzing,” as compared to “the girl is fezzing,” provides more information about which event, out of multiple candidate events, is being labeled; thus, we may expect it to better facilitate verb learning. However, we find evidence to the contrary: in a verb learning study, preschoolers (N = 60, mean age = 38 months) only performed above chance when the subject was an unmodified determiner phase, but not when it was modified (Experiment 1). Experiment 2 replicated this pattern with a different set of stimuli and a wider age range (N = 60, mean age = 45 months). Further, in Experiment 2, we looked at both learning outcomes—by evaluating pointing responses at Test, and also the learning process—by tracking eye gaze during Familiarization. The results suggest that children’s limited processing abilities are to blame for poor learning outcomes, but that a nuanced understanding of how processing affects learning is required.

Acknowledgments

We are grateful to the families who participated in this study, to the members of the Boston University Child Language Lab who assisted with participant recruitment and data collection, and to anonymous reviewers who helped us improve the quality of the manuscript. A subset of the data was published in the Proceedings Supplement of the 40th annual Boston University Conference on Language Development. This research was supported by NIH grants K01DC013306 and R01DC016592. The content is solely the responsibility of the authors and does not necessarily represent the official view of the National Institutes of Health.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ Our use of the term “informativity” differs from that in the pragmatics literature (i.e. the Quantity Maxim, Grice, 1975). According to the Quantity Maxim, the speaker provides as much information as is needed, no more (otherwise “over-informative”) and no less (otherwise “under-informative”). We use “informativity” in a similar sense as the term “information density” in the computational linguistics literature, which describes the amount of information conveyed in an utterance (Aylett & Turk, 2004; Jaeger & Levy, 2007). Thus, an utterance that contains more truthful information is by our definition higher in informativity.

² “Light” and “Heavy” in the condition names distinguish conditions that vary in informativity: “Light” is lower and “Heavy” is higher in informativity. “Necessary” and “unnecessary” distinguish conditions that vary in pragmatic felicity: “necessary” follows the Quantity Maxim (Grice, 1975) but “unnecessary” provides over-informative statements.

³ We did not use wholly novel actions because it is difficult to come up with novel actions that could be plausibly lexicalized by a single verb in a human language, and because there is no evidence that children engage in mutual exclusivity when learning novel verbs as they do for nouns (Hesketh & Ellis Weismer, 1997; Merriman, Marazita, & Jarvis, 1995).

⁴ One parent of a participant in the Light condition did not provide vocabulary data.

⁵ One of the eight adjectives, “round,” was inadvertently omitted from the vocabulary checklist.

⁶ One parent of a participant in the Light condition did not provide vocabulary data.

⁷ We also reported effects with a p-value between 0.05 and 0.1 but did not use them as bases for interpretation.

⁸ The range of performance was 53%-65%, similarly to other verb learning studies using these tasks; Arunachalam and Waxman (2015) and Syrett et al. (2014) each had a 65% success rate in the condition with the best performance.

⁹ This was to reduce confusion caused by the fact that in our setup, the experimenter behind the curtain had a mirror-image perspective of the child’s behavior.

¹⁰ The model failed to converge with a participant-by-condition random effect on both time terms, so this random effect on the quadratic time term was removed.

¹¹ Although there were no statistically significant differences between the Heavy-Necessary condition and the Heavy-Unnecessary condition, trends suggested that the former may have been more difficult. If further work suggests that this is true, this could be because children could not simply “listen through” the modifier (Thorpe & Fernald, 2006) but instead were required to parse and integrate it, and/or because the visual scenes in the Heavy-Necessary condition (e.g., two boys) were more similar to each other than in the Heavy-Unnecessary condition (e.g., boy and girl), potentially making it more difficult to distinguish them (Arias-Trejo & Plunkett, 2010; Bergelson & Aslin, 2017).

Additional information

Funding

This work was supported by the National Institutes of Health [NIH K01DC013306,R01DC016592].