(Apologies for multiple postings and reminders)
Composes end-of-project workshop: Call for participation
Workshop website: http://clic.cimec.unitn.it/composes/workshop.html
The end-of-project workshop of the Composes project
(http://clic.cimec.unitn.it/composes/) will take place on Sunday
August 14th 2016 in Bolzano (Italy), as a satellite event of ESSLLI
2016 (http://esslli2016.unibz.it/).
The workshop will be an occasion to discuss some exciting topics in
computational semantics, with some great invited speakers/panelists
leading the discussion. We foresee a mixture of position statements on
the topics below by the invitees and audience participation in the
form of open debates.
Speakers/Panelists:
- Nicholas Asher
- Marco Baroni
- Stephen Clark
- Emmanuel Dupoux
- Katrin Erk
- Adele Goldberg
- Alessandro Lenci
- Hinrich Schütze
- Jason Weston
Topics:
- Lessons learned from the Composes project: Which problems were we
trying to solve? Have we solved them? Have new-generation neural
networks made compositional distributional semantics obsolete?
- End-to-end models and linguistics: What is the role of linguistics
in the (new) neural network/end-to-end/representation learning era?
Do such systems need linguistics at all? Are some linguistic
theories better tuned to them than others? Is there an appropriate
vocabulary of linguistic units for end-to-end systems? Is
compositionality a solved problem? Which linguistic challenges are
difficult to tackle with neural networks?
- "Fuzzy" vs "precise" (concepts vs entities, generics vs specifics,
lexical vs phrasal/discourse semantics, analogy vs reasoning, sense
vs reference): Are learning-based statistical methods only good at
fuzzy? Can new-generation neural networks (Memory Networks, Stack
RNNs, NTMs etc) handle both fuzzy and precise? Is fuzzy a solved
problem?
- Learning like humans do: If we want to develop systems reaching
human-level language understanding, what is the appropriate input?
What should training data and objective functions look like? What
are appropriate tests of success? Assuming our methods are much more
data-hungry than human learning is, why is this the case? Ideas for
fixing that? What ways can we teach our models to understand, other
than through expensive labeling of data?
Please visit the workshop website for information about (free)
registration and for updates: