Abstract
There is mounting evidence that existing neural network models, in particular the very popular sequence-to-sequence architecture, struggle with compositional generalization, i.e., the ability to systematically generalize to unseen compositions of seen components. In this paper we demonstrate that one of the reasons hindering compositional generalization relates to the representations being entangled. We propose an extension to sequence-to-sequence models which allows us to learn disentangled representations by adaptively re-encoding (at each time step) the source input. Specifically, we condition the source representations on the newly decoded target context which makes it easier for the encoder to exploit specialized information for each prediction rather than capturing all source information in a single forward pass. Experimental results on semantic parsing and machine translation empirically show that our proposal yields more disentangled representations and better generalization.
Abstract (translated)
URL
https://arxiv.org/abs/2110.04655
