panelist: Donia Scott
Title: Multilingual NLG revisited?

Multilingual natural language generation was proposed in the
early-1990s as an alternative to Machine Translation. Instead of
translating from an input text in one language to an (ideally
semantically equivalent and syntactically well-formed) target text in
another language, the aim was to map directly from semantics into
texts in multiple languages. The advantage of the M-NLG approach over
MT was seen to be the improvement in fluency, especially at the levels
of discourse and pragmatics, that could be achieved by avoiding biases
introduced by the input language.  While some notable work in M-NLG
continues, the venture has been largely abandoned by the NLG community
since the early part of the new century, in part because the limits of
scaleability and domain dependence made the method unattractive for
real-world applications and in part because of the increasing success
in statistical MT methods. Within statistical MT, the dominant focus
on the sentence over larger discourse units has, however, done little
to address issues of pragmatic congruence and discourse
equivalence. Are there new opportunities to address these issues? For
example, with the growing popularity of neural MT as a dominant
paradigm can we expect to see multilingual equivalents with improved
fluency beyond the scope of sentences?