Computational models of language acquisition

Question

What are currently used computational models/frameworks of language acquisition?

Desired features:

• models that are biologically plausible (such as neural networks). Personally, I have used cascade correlation neural nets to model pronoun acquisition in children, but in hindsight, I consider this framework to be not neurobiologically plausible.

• models/frameworks that can model learning of both L1 and L2 in children and adults, and capture the observed differences.

Related question:

Computational models of early learning in children on CogSci.SE

Answer 1

I'll break up my response into categories.

Literature Reviews / Surveys: An OK survey of various models in psycholinguistics, cognitive science, and computer science is "Computational Models of Language Acquisition" by Shuly Wintner. Cambridge has a good paper on the Categorial Grammar Learner (CGL) model, which is "Computational models for first language acquisition" by Paula J. Buttery.

They build on previous models (such as the Triggering Learning Algorithm of Gibson and Wexler and the Structural Triggers Learner of Fodor and Sakas) and improve CGL.

Books: A decent book, always "Computational Models of Language Acquisition", has been written on the subject by Wintner which covers several accepted models, and a recent similar book (same title as the other one) has been published on the topic by Charles Yang (but I have never read the most recent work, so I cannot vet for its quality, but it appears to be good).

A Google search for 'computational models of language acquisition' will reveal many more results, but these are some that I've found helpful.

Answer 2

Well, in the 10 years since this question was posed, a lot happened.

Neural networks became the main approach to language models and machine translation, the tasks that best represent language acquisition.

Translation is not language acquisition, but, for humans anyway, translation requires having acquired the languages.

The neurons of a neural network have nothing to do with biology, as far as I can tell.

But plenty of the techniques do happen to model aspects of human language acquisition:

• subword representation, e.g. byte-pair encoding
• learning from monolingual data in the target language, e.g. back-translation
• multilingual models and transfer learning
• generating sequences in order, e.g. decoding

That is, we humans do not acquire a language mainly from parallel data, whether words, phrases or corpora.

Like neural models, we even suffer from hallucination and catastrophic forgetting.