A newly published study on arXiv (ID 2610.31842v1) investigates how children acquire the meanings of a large number of simple, concrete words in early vocabulary, focusing on the theory of “cross-contextual learning”: that is, children infer word meanings by repeatedly accumulating statistical information about the co-occurrence of words and their referents.
The research team conducted experiments on the BabyView corpus using a contrastive vision–language model—the corpus comprises 2,441 hours of longitudinal first-person (egocentric) video from 48 children, simultaneously recording their real-world visual and linguistic inputs. The results show that the model can successfully learn object–label mappings, but its performance is highly dependent on the quality of visual representations: using internet-pretrained visual features yields significantly better results than learning visual representations solely from developmental videos.
The analysis also found that the vast majority of learning signals originate from a small subset of speech segments that explicitly describe visible objects; the word‑meaning representations learned by the model exhibit human‑like semantic structures, yet the trajectory of its learning process as it evolves with training experience differs substantially from that of real children. The paper notes that cross‑contextual learning is a viable strategy for early noun acquisition, but building computational models that align with developmental principles urgently requires overcoming the technical bottleneck of efficiently learning high‑quality visual representations from naturalistic data.