Acceso usuarios
BabyLMs are neural models which aim to simulate child language acquisition by training on developmentally plausible corpora. In this framework, several strategies have been explored to bridge the gap between machine and human language learning, most notably Curriculum Learning - a training approach that presents examples to a neural network in order of increasing linguistic complexity. While numerous ordering criteria have been proposed, this study leverages four distinct metrics - two based on lexical diversity and two on syntactic complexity. We investigate whether exposing models to language in a manner similar to human development yields improvements in the generalization of linguistic abilities.
