Compositional Generalisation with Structured Reordering and Fertility Layers
This addresses a key limitation in neural models for tasks like semantic parsing, though it is incremental by building on prior reordering work.
The paper tackles the problem of compositional generalization in seq2seq models by introducing a neural model with fertility and reordering steps, which outperforms seq2seq models by a wide margin on challenging semantic parsing tasks requiring generalization to longer examples.
Seq2seq models have been shown to struggle with compositional generalisation, i.e. generalising to new and potentially more complex structures than seen during training. Taking inspiration from grammar-based models that excel at compositional generalisation, we present a flexible end-to-end differentiable neural model that composes two structural operations: a fertility step, which we introduce in this work, and a reordering step based on previous work (Wang et al., 2021). To ensure differentiability, we use the expected value of each step. Our model outperforms seq2seq models by a wide margin on challenging compositional splits of realistic semantic parsing tasks that require generalisation to longer examples. It also compares favourably to other models targeting compositional generalisation.