Child-directed speech is optimized for syntax-free semantic inference

Abstract

The way infants learn language is a highly complex adaptive behavior. This behavior chiefly relies on the ability to extract information from the speech they hear and combine it with information from the external environment. Most theories assume that this ability critically hinges on the recognition of at least some syntactic structure. Here, we show that child-directed speech allows for semantic inference without relying on explicit structural information. We simulate the process of semantic inference with machine learning applied to large text collections of two different types of speech, child-directed speech versus adult-directed speech. Taking the core meaning of causality as a test case, we find that in child-directed speech causal meaning can be successfully inferred from simple co-occurrences of neighboring words. By contrast, semantic inference in adult-directed speech fundamentally requires additional access to syntactic structure. These results suggest that child-directed speech is ideally shaped for a learner who has not yet mastered syntactic structure.

How the characteristics of words in child-directed speech differ from adult-directed speech to influence children’s productive vocabularies

“…Third, sentence structure and word content are changed from what would be used when talking to other adults. For example, Furrow et al (1979) showed that speech directed toward 18- to 27-month-old children had greater numbers of wh -questions and fewer declaratives than adult-directed speech, while You et al (2021) showed that child-directed speech allowed causal meaning to be inferred based on the co-occurrence of neighboring words, which was not possible with adult-directed speech.…”

Explainability of neural networks for child language: Agent‐First strategy in comprehension of Korean active transitive construction

“…You et al. (2021), one recent study relevant to the current work, showed that a Word2Vec learner was able to conduct semantic inference from raw caregiver input without the mediation of structural information, by measuring the model's discrimination performance on English causatives. However, other than You et al.…”

“…The continuous development of NN algorithms in computer science has made their internal mechanisms deviate from how biological neurons operate in the human brain (e.g., Crick, 1989), but NNs have been applied to various disciplines in reality (Abiodun et al., 2018). Specifically in the literature on language development, while researchers adopt various computational modelling techniques to reveal developmental trajectories of linguistic knowledge (Alishahi & Stevenson, 2008; Ambridge et al., 2020; Bannard et al., 2009; Chang, 2009; Divjak et al., 2021; You et al., 2021), the current research practice bears two major caveats. First, findings are based exclusively on a limited range of languages such as English, generating a sampling bias towards those languages and populations speaking those languages (cf.…”

“…However, other than You et al. (2021), computational research on child language through NNs is extremely thin.…”

The function and evolution of child-directed communication

“…Repetitions of constructions at the beginning of utterances (e.g., this is an X [ 62 , 90 ]) and discontinuous repetitions (e.g., I X you [ 91 , 92 ]) are ubiquitous and support the generalization of word classes, such as nouns and verbs [ 93 ]. In addition, repetitive structures or distribution of words surrounding specific verbs support the generalization of meaning [ 94 ], and the high number of repetitions found in CDC are positively correlated with word comprehension [ 95 , 96 ]. A specific form of repetitions frequently used in CDC is variation sets, successive utterances with partial self-repetitions produced by caregivers [ 71 , 72 ], which themselves are positively related to better linguistic outcomes in naturalistic longitudinal [ 97 ], and experimental settings [ 98 ].…”