Compositionality---the principle that the meaning of a complex expression is built from the meanings of its constituent parts---is considered a central property of human language. The key benefit of compositionality is compositional generalization, which enables the production and comprehension of novel expressions analyzed as new compositions of familiar parts. In this presentation, I discuss my work on developing a test for compositional generalization for artificial neural networks based on human generalization patterns discussed in existing linguistic and developmental studies, and applying this test to several instantiations of Transformer (Vaswani et al. 2017) and Long Short-Term Memory (Hochreiter & Schmidhuber 1997)) models to better characterize their learning biases.

CSC Speaker Event: Najoung Kim
Date: Thursday (10/21) 6-7pm (ET)
Link: https://umich.zoom.us/j/98135198767