Language Acquisition with Echo State Networks: Towards Unsupervised Learning

The modeling of children language acquisition with robots is a long quest paved with pitfalls. Recently a sentence parsing model learning in cross-situational conditions has been proposed: it learns from the robot visual representations. The model, based on random recurrent neural networks (i.e. reservoirs), can achieve significant performance after few hundreds of training examples, more quickly that what a theoretical model could do. In this study, we investigate the developmental plausibility of such model: (i) if it can learn to generalize from single-object sentence to double-object sentence; (ii) if it can use more plausible representations: (ii.a) inputs as sequence of phonemes (instead of words) and (ii.b) outputs fully independent from sentence structure (in order to enable purely unsupervised cross-situational learning). Interestingly, tasks (i) and (ii.a) are solved in a straightforward fashion, whereas task (ii.b) suggest that that learning with tensor representations is a more difficult task