Language Background, Semantic Complexity and Statistical Learning Aptitude in Classifier Learning: Evidence from Human and MLLM

The process of language acquisition is driven by an intricate interplay of linguistic, cognitive, input and socioemotional factors. However, the specific contributions and operational mechanisms of these factors remain largely elusive. Particularly, learning classifiers, an important linguistic marker in Chinese that obligatorily connects a numeral and a noun, has long been considered difficult to acquire. The study examined how classifier learning is modulated by language background, semantic complexity and visual statistical learning aptitude (VSL) via a pseudo-classifier learning task which controlled confounding variables of input frequency. The performance of human learners and multimodal large language models (MLLMs) was also compared. Seventy-eight adults with different language backgrounds (Mandarin, Cantonese, and non-classifier language speakers, NCL) and ten state-of-the-art MLLMs implicitly learned two pseudo-classifiers in the pseudo-classifier learning task with invented novel objects. The two classifiers differed in semantic complexity: a shape classifier encoding a language-specific shape feature (denoting fluffy and curly-haired objects), and a collective classifier encoding a conceptually universal collective feature (denoting objects in triplets with no fluffy or curly-haired features). Each participant and model received 36 exposures for each classifier and was tested on their production and comprehension of classifier. Human participants had an additional VSL aptitude test. Results showed that three factors under investigation did not independently modulate classifier learning, but jointly shaped the learning process. The effect of language background interacted with semantic complexity, as the collective classifier encoding a conceptually universal feature was easier to acquire for NCL, supporting the unified competition model framework. Similarly, VSL aptitude interacted with language background and task difficulties: VSL aptitude mattered most when language-specific prior knowledge was absent and when the task was demanding. Significant discrepancies were also found between machine and human learners, which could not be bridged simply through model scaling or an extended in-context window.