I was very impressed by a paper we recently read in our reading group, which showed that small differences in BLEU scores for MT usually dont mean anything. Since lots of academic papers justify a new model on the basis of such small differences, this is a real problem for NLP.
NLP technology has changed and advanced over the past two decades, but it often seems that NLG evaluation has not. Why is the 18-year old BLEU metric still so dominant?
I was recently asked by someone if it was possible to easily determine whether an NLP system was good enough for a specific use case. Currently this is very hard. Making it easy could be a “grand challenge” for evaluation!
Many neural NLG systems “hallucinate” non-existent or incorrect content. This is a major problem, since such hallucination is unacceptable in many (most?) NLG use cases. Also BLEU and related metrics do not detect hallucination well, so researchers who rely on such metrics may be misled about the quality of their system.
My advice on how to perform a high-quality validation study, which assesses whether a metric (such as BLEU) correlates well with human evaluations.
BLEU works much better for MT systems and NLG systems. In this blog I present some speculations as to why this is the case.
My structured survey of BLEU suggests that BLEU-human correlations are worse in German than in many other languages. But there are many caveats, so we need to be cautious in interpreting this result.
The correlation between BLEU and human evaluations of MT systems seems to be increasing over time. Since BLEU has not changed, how is this possible, and what does it mean?
The first phase of my systematic review of BLEU shows that BLEU-human correlations are all over the place, and that none of the studies in my review have correlated BLEU with real-world utility or user satisfaction.
I’m planning to do a systematic review of the validity of BLEU, and am very keen to get comments and suggestions on study design from others!
