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

Wang Chaojun (王朝君)

[0]

Sennrich Rico

[0]

ACL, pp. 3544-3552, 2020.

Cited by: 7|Views49

Keywords:

Engineering and Physical Sciences Research Councilexposure biasstatistical machine translationerror propagationsequence level trainingMore(11+)

Weibo:

We find that Minimum Risk Training, which does not suffer from exposure bias, can be useful even when it does not increase performance on an in-domain test set: it increases performance under domain shift, reduces the number of hallucinations substantially, and makes beam search ...

Abstract:

The standard training algorithm in neural machine translation (NMT) suffers from exposure bias, and alternative algorithms have been proposed to mitigate this. However, the practical impact of exposure bias is under debate. In this paper, we link exposure bias to another well-known problem in NMT, namely the tendency to generate halluci...More

Code:

https://github.com/zippotju/Exposure-Bias-Hallucination-Domain-Shiftmd

Data:

Full Text

View via Publisher