Fixing exposure bias with imitation learning needs powerful oracles

This addresses exposure bias in neural machine translation for NLP researchers, but appears incremental as it evaluates a specific oracle limitation.

The paper tackled the neural machine translation exposure bias problem using imitation learning with error-correcting oracles, finding that an SMT lattice-based oracle performed well in translation tasks but was too pruned and idiosyncratic to effectively serve as the oracle for imitation learning.