Contextual inferences over speakers' pragmatic intentions: Preschoolers' comprehension of contrastive prosody

Contextual inferences over speakers’ pragmatic intentions: Preschoolers’ comprehension of contrastive prosody Chigusa Kurumada kurumada@stanford.edu Department of Linguistics Stanford University Abstract Speaker intention We investigate pre-schoolers’ ability in drawing pragmatic in- ferences based on prosodic information. Previous work has found that young children are generally oblivious to intona- tional meaning of utterances. In particular, the ability to com- prehend contrastive prosody develops late during language ac- quisition (after the age of 6). In three experiments, we show that preschoolers can engage in prosody-based pragmatic in- ferences if the context provides supports for them. Further- more, we find that preschoolers’ interpretation of prosody in- volves complex counter-factual reasoning (‘what the speaker would have said if she had intended another meaning’). The picture emerging from our studies contrasts with previous work: Through rich contextual inferences, four-year olds are able to bootstrap their interpretation of prosodic information, and achieve adult like performance in intonation interpretation. Keywords: Prosody, language acquisition, contrastive accent, Principle of Contrast, rational inference Intention to express meaning m1 I ntention to express meaning m2 Inference Linguistic signal signal s1 signal s2 Figure 1: A schematic representation of a pragmatic model of intention-signal mapping In the model illustrated in Figure 1, a speaker’s intention (to express meaning m1) generates a particular linguistic sig- nal (signal s1); If the speaker had meant to express meaning m2, she would have generated a different signal (signal s2). A word learner’s job is to work backwards from the observed signal to infer the speaker’s intention (as indicated with ar- rows) while updating her belief about the signal-intention mappings, including any associations newly introduced. The current study extends this idea to a new domain: Young children’s interpretation of contrastive prosody. Just as words are associated with speaker intentions, different prosodic representations are probabilistically mapped onto intentions that had generated them (Figure 2). Compared to words, however, prosodic signals are continuous and variable, which can make the mapping puzzle much harder to solve. Furthermore, the intentions that prosodic representations en- code are often very abstract (e.g., contrastiveness) and are not always disambiguated in an observable context. In other words, signal-intention mappings for prosody include much more uncertainty (indicated by the thinner arrows in Figure 2) for listeners to overcome. The current study suggests that, despite this additional complexity, the rational inferences at- tested in word-learning provide leverage in young children’s discovery of pragmatic functions of contrastive prosody as well. Prosodic information is known to encode structural boundaries and phrasing but also speakers’ intentions (Pierrehumbert & Hirschberg, 1990; Ladd, 2008; B¨uring, 2003). Much attention has been paid to how listeners interpret context-relevant contrast based on a low-high-low (an L+H*) pitch accent. (e.g., KATIE (L+H*) did not win a truck (but LAURA did); Ito & Speer, 2008). Previous work has gener- ally agreed that inferences based on an L+H* accent present great difficulty to preschoolers, and even young school chil- dren fail to achieve adult-like performance in experimental settings (e.g., Solan, 1980; Cruttenden, 1985; Wells, Peppe, Introduction In learning new words, young children can make use of pragmatic inference to bootstrap their knowledge about new word-object mapping. For example, in a situation where an adult utters “Give me the TOMA (nonce word)” 1 when a fa- miliar object (e.g., a spoon) and an unfamiliar object (e.g., a whisk) are present in a visual field, a child as young as two years of age is likely to reach for the whisk. This is con- sidered to be based on a cognitive bias for a unique object- label mapping (e.g., Markman and Wachtel (1988)) or the in- ference that the mother should have used the familiar word (spoon) if she had intended to refer to it (e.g., Clark (1990)). Such pragmatic dispositions provide immense leverage in word learning because there is inherent uncertainty associ- ated with mappings between speakers’ intentions, linguistic signals, and their referents (Frank, Goodman, & Tenenbaum, 2009). One way to systematically solve this puzzle is to esti- mate the probability assigned to a possible intention-signal mapping relative to other possible mappings warranted by the same context. For instance, the probability of the sig- nal “Give me the TOMA” expressing the speaker’s intention of picking out the non-spoon object is estimated in proportion to probabilities of (1) the signal being generated by the inten- tion of picking out a spoon; and (2) other signals (including [a spoon]) being generated by the intention of picking out the target (non-spoon) object. 1 Hereafter: double-quotation marks are used for quoting speech, with phonetic and prosodic specification. Capital letters represent prosodic emphasis. Square brackets ([ ]) are used for example words or sentences abstracted away from acoustic detail, e.g., [It is raining outside] can be said as, “It’s RAINING outside!”, “It IS raining out- side!”, et cetera.)