Early Birds Can Fly: Awakening the Literal Meaning of Conventional Metaphors Further Downstream

Abstract

ABSTRACTConventional metaphors such as early bird are interpreted rather fast and efficiently. This is so because they might be stored as lexicalized, non-compositional expressions. In a previous study, employing a maze task, we showed that, after reading metaphors (John is an early bird so he can …), participants took longer and were less accurate in selecting the appropriate word (attend) when it was paired with a literally-related distractor (fly) rather than an unrelated one (cry). This suggests that the literal meaning of conventional metaphors is awakened or made available immediately after their metaphorical interpretation. But does the literal meaning remain available further downstream during sentence comprehension? In two experiments also employing a maze task, we examined whether the awakening effect can be obtained when there is a medium (6 to 8 words) and a large (11 to 13 words) distance between the metaphor and lexical choice. Results indicated that the metaphor awakening effect persists but decreases as word distance increases. An analysis of our data based on a GPT model showed that our maze effects could not be attributed to target predictability. Overall, our results suggest that the literal meaning of a metaphor is accessed and remains available for about three seconds, fading as the sentence unfolds over time. The results support a model of metaphor comprehension that postulates the availability of both literal and metaphoric content in the course of sentence processing. AcknowledgmentsWe are indebted to Tobias Ungerer for his comments on section 5 and the calculation of the surprisal scores using the GPT-2 model, and to Cedric Le-Bouar for helping code the data. We thank Caitlyn Antal for her guidance on the statistical analyses for section 5. We also thank the two anonymous reviewers for their invaluable comments on an earlier version of the present article.Disclosure statementNo potential conflict of interest was reported by the author(s).Notes1 We note, however, that the present study was not designed to investigate the conceptual metaphor theory (Lakoff, Citation1993; Lakoff & Johnson, Citation1980), nor were our materials suited for such a task. Although some of the cues we employed may be taken as novel extensions of conventional metaphors, we cannot assert that (a) all of our cues are metaphorical extensions, as some may be literal cues. For instance, we anticipate that the literal meaning of cold feet can be triggered by the cue warm regardless of whether the latter is used metaphorically (e.g., warm welcome) or literally (e.g., warm weather). Nor can we assert that (b) all extensions belong to the same metaphor family, for we may not have the theoretical grounds to establish all cases in which a metaphor belongs to one or another metaphor family. For instance, it is not obvious whether warm blood, warm gesture, hot take, hot minute, cold glance, cold turkey, cool cat, and cool head belong to the same metaphor family.2 We employed a more conservative approach following Forster, Guerrera, and Elliot (Citation2009) procedures (i.e., removing RTs longer than 1500 ms and replacing remaining outliers per participants with 2 SD above the mean).3 A language model such as GPT−2 can perform language tasks such as reading comprehension, summarization, translation, and question answering. In addition, the GPT−2 model yields reliable estimates of cloze probabilities as it has been trained on approximately 8 M webpages to predict the next word given the previous ones (Radford et al., Citation2019). To obtain our surprisal scores, we used the large version of the GPT−2 model, which contains 762 M parameters and 36 layers (Radford et al., Citation2019).4 We note that single-word recognition times are in the order of milliseconds, with some classical RSVP studies suggesting that, with about 60 ms of exposure, words can be recognized and integrated into an ongoing propositional representation of the sentence (see, e.g., Forster, Citation1970; Potter, Citation2018, for a review).Additional informationFundingThis research was supported by grants from the National Science and Engineering Research Council (NSERC) and the Social Sciences and Humanities Research Council (SSHRC) to RGdA, and by a Doctoral Fellowship from the Fonds de Recherche du Québec - Société et Culture (FRQSC) to LP.