From Kevin Duh:

I wonder if the issue of getting better translation probabilities for
phrase pairs is an important issue in phrase-based SMT? The paper
presents an interesting solution, but are there more work out there? I
see an analogue to the problem of data sparsity for ngram language
modeling using long contexts--in ngram modeling, we also have this
problem of unrobust probabilities when using ngrams with many
conditioning factors. As a result, we have a whole slew of research
(e.g. smoothing, class-based LM, etc.) that tries to address this
problem. I wonder if we can modify some of the LM techniques to deal
with phrase pair probabilities? (For example, Katrin suggested using
FLMs to gain robust estimates through morphology.)