How Words Can and Cannot Be Learned by Observation

Abstract

Three experiments explored how words are learned from hearing them across contexts. Adults watched 40-s videotaped vignettes of parents uttering target words (in sentences) to their infants. Videos were muted except for a beep or nonsense word inserted where each “mystery word” was uttered. Participants were to identify the word. Exp 1 demonstrated that most (90%) of these natural learning instances are quite uninformative, whereas a small minority (7%) are highly informative, as indexed by participants’ identification accuracy. Preschoolers showed similar information sensitivity in a shorter experimental version. Two further experiments explored how cross-situational information helps, by manipulating the serial ordering of highly informative vignettes in five contexts. Response patterns revealed a learning procedure in which only a single meaning is hypothesized and retained across learning instances, unless disconfirmed. Neither alternative hypothesized meanings nor details of past learning situations were retained. These findings challenge current models of cross-situational learning. Learners appear to use a one-trial fast mapping procedure, even under conditions of referential uncertainty.

Subjective confidence influences word learning in a cross-situational statistical learning task

“…A mixed-effect model with random intercept for each participant showed a significant interaction of chosen Object Type (target vs. competitor) and Condition (Early vs. Late) on confidence scores (χ 2 (1) = 8.42; p = .004; Cohen's d = 0.70; no other effect was significant; see Appendix Such order effects have also been previously reported elsewhere (Medina et al, 2011;Thaker, Tenenbaum, & Gershman, 2017), and have been accounted for by participants' past selection history consistent with an hypothesis-testing account (K. Smith et al, 2011;Stevens, Gleitman, Trueswell, & Yang, 2017;Thaker et al, 2017;Trueswell et al, 2013;Yurovsky & Frank, 2015). Here, however, we suggest that this is a result of participants being more likely to display a confirmation bias in the late informative condition when they hold high confidence in the competitor object.…”

“…Frank, Goldwater, & Keller, 2009;Siskind, 1996;. Other evidence supports an hypothesis-testing mechanism in which learning is more discrete, with learners selecting the most likely meaning for a word in a given moment and subsequently confirming or falsifying this hypothesis as new information becomes available in subsequent word usages (Medina, Snedeker, Trueswell, & Gleitman, 2011;Trueswell, Medina, Hafri, & Gleitman, 2013;Yang, 2020). While the use of one or the other mechanism may depend on attentional and memory demands (Yurovsky & Frank, 2015), both mechanisms focus on how learners use their objective experience with the world, in and across learning exposures, to generate and evaluate word meaning hypotheses and do not attempt to capture the influence of more subjective processes.…”

Cross-Situational Learning Is Supported by Propose-but-Verify Hypothesis Testing

“…(F) Plots displaying % signal change estimates during the learning trials and change in performance statistics for the regions identified in (E). tions are carried forward until they are verified or disconfirmed on later encounters (e.g., [6]). The strongest version of a local model is propose-but-verify (PbV) hypothesis testing (see [7]).…”

“…Nonetheless, because names for concrete nouns are constant, we are able to learn them across successive encounters [2,3]. This form of ''cross-situational'' learning may result from either associative mechanisms that gradually accumulate evidence for each word-object association [4,5] or rapid proposebut-verify (PbV) mechanisms where only one hypothesized referent is stored for each word, which is either subsequently verified or rejected [6,7]. Using model-based representation similarity analyses of fMRI data acquired during learning, we find evidence for learning mediated by a PbV mechanism.…”

The temporal structure of naming events differentially affects children’s and adults’ cross-situational word learning

“…Further, word learning events are often ambiguous. Children may hear words without an ostensive cue or clear context to suggest the exact referent for the label (Medina, Snedeker, Trueswell, & Gleitman, 2011;Quine, 1960). Thus, the key to successful word learning is resolving referential ambiguity across multiple word learning events that are distributed in time.…”