English

Fixing exposure bias with imitation learning needs powerful oracles

Computation and Language 2021-09-20 v2 Machine Learning

Abstract

We apply imitation learning (IL) to tackle the NMT exposure bias problem with error-correcting oracles, and evaluate an SMT lattice-based oracle which, despite its excellent performance in an unconstrained oracle translation task, turned out to be too pruned and idiosyncratic to serve as the oracle for IL.

Cite

@article{arxiv.2109.04114,
  title  = {Fixing exposure bias with imitation learning needs powerful oracles},
  author = {Luca Hormann and Artem Sokolov},
  journal= {arXiv preprint arXiv:2109.04114},
  year   = {2021}
}
R2 v1 2026-06-24T05:49:00.239Z