Revisiting Iterative Back-Translation from the Perspective of Compositional Generalization

Human intelligence exhibits compositional generalization (i.e., the capacity to understand and produce unseen combinations of seen components), but current neural seq2seq models lack such ability. In this paper, we revisit iterative back-translation, a simple yet effective semi-supervised method, investigate whether how it can improve generalization. work: (1) We first empirically show that back-translation substantially improves performance on benchmarks (CFQ SCAN). (2) To why is useful, carefully examine gains find increasingly correct errors in pseudo-parallel data. (3) further encourage mechanism, propose curriculum which better quality data, thus improving performance.