Recurrent babbling: evaluating the acquisition of grammar from limited input data

TL;DR

This paper investigates how Long Short-Term Memory networks (LSTMs) can learn grammatical structures from realistic, limited child-directed input, demonstrating that they develop abstraction over time similar to language acquisition.

Contribution

It introduces a novel methodology to quantify grammatical abstraction in neural network output and shows LSTMs can learn and generalize grammar from limited, realistic data.

Findings

LSTMs abstract grammatical structures over time

The methodology quantifies grammatical development in neural models

LSTMs can learn from realistic, limited input data

Abstract

Recurrent Neural Networks (RNNs) have been shown to capture various aspects of syntax from raw linguistic input. In most previous experiments, however, learning happens over unrealistic corpora, which do not reflect the type and amount of data a child would be exposed to. This paper remedies this state of affairs by training a Long Short-Term Memory network (LSTM) over a realistically sized subset of child-directed input. The behaviour of the network is analysed over time using a novel methodology which consists in quantifying the level of grammatical abstraction in the model's generated output (its "babbling"), compared to the language it has been exposed to. We show that the LSTM indeed abstracts new structuresas learning proceeds.