Maria Spychalska, Ludmila Reimer and Peter Hagoort
Affiliation: Max Planck Institute for Psycholinguistics, Ruhr University Bochum, Max Planck Institute for Psycholinguistics
Category: Linguistics
Keywords: implicature, perspective-taking, EEG, Virtual reality, N400
Date: Tuesday 2nd of September
Time: 15:00
Location: GSSR Plenary Hall (268)
View the full session: Quantifiers, Plurals & Numbers
In everyday communication, we do not only take utterances at face value, but we need to interpret the speaker's meaning, which involves considering their communicative intentions and grasping their epistemic perspective, that is, information access. Consider a card game where one (knowledgeable) player can check all cards dealt in the round, while the other (ignorant) player can only see a subset. It would be odd if the ignorant player stated “all cards contain cats” because the truth or falsity of the sentence with “all” is only accessible to the knowledgeable player. By contrast, sentences like “some cards present cats” are more expected from the ignorant player. They are, however, pragmatically ambiguous due to conveying the scalar implicature “not all”. In situations where all cards are known to contain cats, “some cards contain cats” is considered underinformative, whereas the more informative alternative with “all” is more appropriate from a cooperative speaker (Maxim of Quantity, Grice (1989)). Uttered by an ignorant speaker, “some” implies that the speaker does not have sufficient information to use the more informative quantifier “all” (weak implicature, Geurts (2010); Sauerland (2004)). Alternative, grammatical accounts (Chierchia et al., 2011) predict that scalar implicatures can be derived by local strengthening of the weak scalar terms. Thus, the listener may expect the speaker not to utter “some” in situations in which they don’t know whether the implicature holds (strong implicature). If communication requires considering the epistemic perspective of our interlocutor in interpreting their utterance, the question arises: Do listeners integrate the speaker's perspective incrementally (take the speaker's perspective), or is our linguistic understanding more egocentric, at least at the early stage of the comprehension process? This question, known as “perspective-taking” problem, has often been addressed in the context of reference resolution, where experiments have brought mixed results, showing both the egocentric bias in the processing as well as an early integration of the speaker’s perspective (Epley et al., 2004; Hanna et al., 2003; Keysar et al., 2000; Keysar et al., 2003; Nadig & Sedivy, 2002; Sedivy, 2003). While existing evidence suggests that listeners/readers consider the speaker’s epistemic perspective in deriving implicatures (Breheny et al., 2013; Dieuleveut et al., 2019; Goodman & Stuhlmüller, 2013; Spychalska et al., 2021), the role of perspective-taking in implicature comprehension is still controversial (Katsos, 2021; Katsos et al., 2023). In an EEG-VR study, we investigate whether the listener integrates the speaker’s perspective while processing unambiguous sentences with “all” and their weaker scalar alternatives with “some”. In our paradigm (Fig.1), participants are listeners and have privileged information in comparison to a virtual speaker (avatar). They were seated in a virtual reality cave, facing a screen displaying four cards and a table with two additional cards. While both the avatar and the participant could see the cards on the screen, the avatar could only see the backside of the table cards, unless they were turned towards her. This setting differs from previous experiments, e.g., Spychalska et al. (2021); Wilson et al. (2023), where participants needed to learn explicit rules about which cards are visible to whom in an abstract screen-placement, and provides a more naturalistic scenario representing visibility of information via its space. The participants were told that they participate in an AI - language training. The avatar described card game situations using sentences such as “All/Some cards contain cats” (fillers used other quantifiers), referring to the displayed cards. While the AI already mastered all the nouns, she still struggled to correctly use quantifiers, especially when missing information about the whole scenario. The participants were asked to give her feedback and help her to learn a natural use of quantifiers. They were instructed to use “some” pragmatically, namely, that “all” would be more natural than “some” in a situation where all cards contain cats and the avatar can see that. However, we did not instruct them to accept or reject “some” in partial information situations, where “all” was possibly true from the avatar’s perspective. Thus, they were free to adopt either the weak or strong implicature interpretation. In all target conditions, only the participants saw the two table cards, creating a partial information situation for the speaker. Each scenario presented two alternative objects. In the Listener-biased context, the two alternatives differed from the listener’s perspective, so that a sentence referring to one alternative was true (for “all”)/pragmatically infelicitous (for “some”), while the sentence referring to the other alternative was false/felicitous, respectively. In the speaker-biased context, the two alternatives differed from the speaker’s perspective, but not the listener’s: One alternative was known to be false (“all”) or pragmatically felicitous (“some”), the other alternative had unknown truth-value/implicature. Filler trials were used, where the table cards turned towards the avatar, so that these cards were not strategically ignored by the listener. Behaviorally, the participants mostly integrated the speaker’s perspective and provided correct feedback, judging “all”-statements as bad if the avatar could not see all cards, irrespective of the actual truth-value of the sentence. Regarding “some”, for full information filler trials, all participants rejected “some” if the implicature was violated, thus, correctly adopting the pragmatic interpretation. However, for target trials, where the avatar had partial information, they mostly accepted “some” statements as good statements (weak implicature). The ERPs results support the egocentric bias hypothesis (see Fig.2 for details). For the unambiguous “all”-sentences, in the listener-biased context, there was an N400 effect for false relative to true statements. In the speaker-biased context, the contrast between the two alternatives was diminished, indicating that the speaker-perspective was not integrated into the meaning-level processing. Interestingly, no effects were observed for “some”-sentences. While the filler trials indicated that the participants calculated the implicature on the behavioral level, the ERPs recorded for the partial information conditions were not modulated by the implicature status or the speaker’s perspective. This suggests that the implicature was possibly inferred at the later stage of sentence processing. Overall, our results show that sentence truth-value was processed egocentrically, while the speaker’s perspective was not integrated during comprehension and did not modulate meaning-related predictive processes.