Understanding Human Statistical Learning in Language Acquisition: Insights from Neural Networks

This project explores how the cognitive mechanisms associated with human statistical learning in language acquisition align with computational processes in three kinds of neural networks: feedforward networks (FFN), simple recurrent networks (SRN), and long short-term memory (LSTM) recurrent networks. Prior research in infants has provided evidence of statistical learning in discovering word boundaries within continuous spoken speech. Replicating statistical learning tasks using neural networks could allow for a better understanding of the fundamentals of these parallel processes in our brains and neural networks alike. This project tested the ability of FFNs, SRNs, and LSTMs to make syllable-by-syllable predictions from sequential data in order to determine if the network could accurately attune to word-like structures. Preference for words over part-words and non-words was measured to see if the network could understand transitional probabilities in the same way that human infants can. The results showed that all three networks could perform as well or better than infants on the same word segmentation tasks, where the LSTM was able to achieve the highest proportion better values for both non-word and part-word tasks, followed by the SRN, and finally the FFN. These results suggest that neural networks, specifically the LSTM, develop internal structures that could behave analogously to the cognitive mechanisms behind human statistical learning in human language acquisition. Additionally, they may provide a foundation for continuing work where I will investigate the limits of each network using more complex learning paradigms.