2018 Friday Poster 6667

Friday, November 2, 2018 | Poster Session I, Metcalf Small | 3pm

Using developmental modeling to specify learning and representation of the passive in English children
E. Nguyen, L. Pearl

Nguyen & Pearl (2017) (N&P) found that verbs with specific combinations of lexical features (termed lexical profiles) consistently have their English verbal be-passive use acquired by children before verbs with other lexical profiles. Using Bayesian modeling, we attempt to capture five-year-old English passivization behavior via a probabilistic integration of lexical feature frequencies in children’s input.

Modeling passivization behavior. From N&P, we identify five lexical profiles that five-year-olds passivize (Table 1). We model a five-year-old’s decision about whether a verb from a specific profile should be passivized as a classification problem: given the profile lexical feature values, should that profile be part of the class of passivizable profiles (c+pass) or not (c−pass)? We use the same lexical features

as N&P (7 total: see Table 2), plus a syntactic feature (TRANSitivity). The reasoning process is implemented via Bayesian inference (1), where each lexical profile feature fi  ∈ F has a particular value vfi  ∈{0, 1}. To determine if a lexical profile with a particular collection of feature values should be in the c+pass class, the modeled learner calculates the posterior probability, based on: (i) the likelihood that a profile comprised of those particular features is c+pass, and (ii) the prior probability of verbs generally being passivizable. To calculate the likelihood, we assess the likelihood of individual feature values (Table 2) using a corpus of child-directed speech from the CHILDES Treebank [2].

We define what the prior would need to be in order for five-year-olds to passivize the verbs they do, assuming they were learning from lexical feature frequency in their input. We attempt to converge on an estimate for the prior on passivization that generates that desired output behavior when that prior and the likelihood are combined. This can define how costly five-year-olds would view passivization to be as a linguistic structure, irrespective of which verbs it applies to. To calculate the necessary c+pass prior, we can compare the likelihoods of c+pass and c−pass (2). In particular, the desired behavior of passivizing a verb with a specific lexical profile (p (c+pass |vf1  … vfn) > 0.50) results when c+pass’s prior ⋅ likelihood > c−pass’s prior ⋅ likelihood.

Results. We looked for c+pass priors < 0.50 as a reasonable estimate, indicating an initial bias against the passive structure. When all features are heeded (Table 3), only verbs from profiles 1 and 2 have a prior like this. This suggests that if children attend to lexical feature frequency, they must be selectively attending to the available lexical features [3].

With a modeled child who selectively attends to one or more lexical features, we’ve so far identified two cases that heeding certain lexical features can yield five-year-old passivization behavior: attending to the TRANS feature exclusively, and attending to the TRANS and OBJ-EXP features only. In both cases, the required c+pass prior < 0.50 for all five verb profiles. This contrasts with a child who attends to the TRANS, OBJ-EXP, and SUBJ-EXP features, who is subsequently unwilling to passivize profile 4 and 5 verbs without a pre-existing general bias in favor of passivization.

Discussion. This is evidence for (i) the utility of selective attention to lexical semantic features, and (ii) five-year-old children viewing the passive structure as somewhat – but not strongly –costly. Future work can determine (i) which other combinations of the 8 available features five-year-olds should attend to in order to yield their passivization behavior, and (ii) apply the same process to three- and four-year-old behavior to yield a quantified snapshot of the developmental trajectory of the passive.

References

  1. Emma Nguyen and Lisa Pearl. 2017. Do you really mean it? Linking lexical semantic profiles and the age of acquisition for the English passive. In The proceedings of the West Coast Conference on Formal Linguistics (WCCFL).
  2. Lisa Pearl and Jon Sprouse. 2013. Syntactic islands and learning biases: Combining experimental syntax and computational modeling to investigate the language acquisition problem. Language Acquisition 20(1):23–68.
  3. Annie Gagliardi, Naomi H Feldman, and Jeffrey Lidz. 2017. Modeling statistical insensitivity: Sources of suboptimal behavior. Cognitive Science 41(1):188–217.