Dave  🦔 is spooky at #INLG2019  🎃

Koehn: Used to be that expanding the beam for SMT yielded consistent improvements at the cost of more compute time. However if you expand the beam too much you start seeing clear degradations in NMT quality

Koehn: We explored the impact of noise on NMT compared to SMT and found across a variety of kinds of noise that NMT is less robust to noise than SMT.

Koehn: To address these problems we created the WMT 2019 filtering task, where the goal is to filter out noisy sentence pairs from the data before training. Looking at systems trained on different amounts of noise we see the same trend across a variety of systems: SMT > NMT

Koehn: What is happening? Why is this happening? Visualization, probing internal states, and trying to trace back decisions to inputs are approaches we are taking to try to understand the problem.

Koehn: One analysis looking to detect hallucinations, we looked at KL-divergence between the predictions of the NMT model for the next word and the LM. High KL-divergence indicates instances where the input matters more than the LM. This approach did not really succeed :/

Koehn: So the next thing we tried was based on the idea of saliency from vision. Idea - if changes in input cause changes in output, then the input mattered. (makes me wonder how the size of the differences between the input and output relate to one another)

Koehn: "Just looking at attention weights often doesn't tell you much", based on Ding et al. 2019 @ WMT, where they showed saliency was more informative wrt word alignments

Koehn: Where to next? We could look at this as an ML problem, arguing that it's overfitting on the data, failing to generalize to data outside the model's comfort zone. Or perhaps arguing that it's 'exposure bias' - the models never see bizarre data during training time...

... so then it doesn't know how to cope with this scenario at test time if it starts a sentence with an odd prefix.

Koehn: So why don't we make the data more diverse? We can do data synthesis (e.g. by vocabulary replacement). I watered my flowers -> I/We/You watered my roses/plants/etc.

Koehn: We can also look at paraphrasing, paraphrasing source, target, or both, and creating new training instances this way.

Koehn: Paraphrasing helps more when you paraphrase the target side (Hu et al. 2019 @ ACL, ongoing work)

Koehn: So far we've been talking about changing data, which is nice, but we can also change how we do ML. For example, adding noise is now a standard practice (e.g. dropout and label smoothing)

Koehn: So what's the outlook? Open Questions - why are neural models overfitting? - can we detect hallucinations? - can we counter-act this behavior? with diversity in training data? with better machine learning?

QA Discussion: what are your thoughts on pipelines? can this help with some of these generalization issues? these days the trend is in the other direction: folks even want to do audio-to-audio MT. so far the syntax-/phrase-based approaches we used to use have not helped with NMT

QA Discussion: detecting hallucinations in target sentences, have you thought about quality in the input? usually we have to translat what we have to translate

QA Discussion: one diff between NLG and MT is that we're starting w/structured meanings as opposed to text-to-text where the meaning may be more ambiguous. the parallel would MT w/back translation similarity to source text. some folks have looked at this...

...but usually after realization is complete rather than online, recognizing where things go wrong as they go wrong