Fixing exposure bias with imitation learning needs powerful oracles

09/09/2021

∙

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.

READ FULL TEXT

Fixing exposure bias with imitation learning needs powerful oracles

Why Exposure Bias Matters: An Imitation Learning Perspective of Error Accumulation in Language Generation

An Empirical Comparison on Imitation Learning and Reinforcement Learning for Paraphrase Generation

Learning Coupled Policies for Simultaneous Machine Translation

Addressing reward bias in Adversarial Imitation Learning with neutral reward functions

Learning to Gather Information via Imitation

On Exposure Bias, Hallucination and Domain Shift in Neural Machine Translation

Learning Neural Parsers with Deterministic Differentiable Imitation Learning

Fixing exposure bias with imitation learning needs powerful oracles

Related Research

Why Exposure Bias Matters: An Imitation Learning Perspective of Error Accumulation in Language Generation

An Empirical Comparison on Imitation Learning and Reinforcement Learning for Paraphrase Generation

Learning Coupled Policies for Simultaneous Machine Translation

Addressing reward bias in Adversarial Imitation Learning with neutral reward functions

Learning to Gather Information via Imitation

On Exposure Bias, Hallucination and Domain Shift in Neural Machine Translation

Learning Neural Parsers with Deterministic Differentiable Imitation Learning