Jump to better conclusions: SCAN both left and right
This addresses dataset design issues for researchers evaluating systematic generalization in sequence-to-sequence models, though it is incremental as it builds on existing SCAN work.
The authors identified limitations in the SCAN dataset for testing systematic generalization in sequence-to-sequence models and proposed a complementary dataset called NACS that better aligns with realistic use-cases, showing that models performing well on SCAN do not necessarily perform well on NACS.
Lake and Baroni (2018) recently introduced the SCAN data set, which consists of simple commands paired with action sequences and is intended to test the strong generalization abilities of recurrent sequence-to-sequence models. Their initial experiments suggested that such models may fail because they lack the ability to extract systematic rules. Here, we take a closer look at SCAN and show that it does not always capture the kind of generalization that it was designed for. To mitigate this we propose a complementary dataset, which requires mapping actions back to the original commands, called NACS. We show that models that do well on SCAN do not necessarily do well on NACS, and that NACS exhibits properties more closely aligned with realistic use-cases for sequence-to-sequence models.