Title:
The Case for Typologically-informed Meaning Representations


Abstract:

Meaning representations, by their very nature, aim to abstract away
from surface forms (which are unpredictably different between
languages) and expose a common level of representation. While this
goal is still out of reach, designing meaning representations that can
be applied consistently across languages is receiving much renewed
attention. I will argue that it is advantageous to consider the
results of descriptive typological research when designing meaning
representation schemes: descriptive typological frameworks can provide
the conceptual backbone and lead to design decisions that are
motivated by evidence from many languages, without constraining the
scheme to a particular theory of the syntax-semantics interface. As an
example of bringing typology into meaning representation for NLP, I
will present UCCA, a broad-coverage semantic representation scheme
that emphasizes cross-linguistic applicability by building on the
widely-used typological framework Basic Linguistic Theory by R.M.W.
Dixon. I will briefly present UCCA, its application to texts in a
number of languages, and its relation with Basic Linguistic Theory. In
ongoing work using UCCA and Universal Dependencies, we are empirically
identifying and quantifying translation divergences and using them to
inform machine translation.

Joint work with Zohar Aizenbud, Leshem Choshen, Daniel Hershcovich,
Dmitry Nikolaev, Jakob Prange, Nathan Schneider, Elior Sulem and Ari
Rappoport.



Brief bio:

Omri is a faculty member in the Hebrew University's departments of
Computer Science and Cognitive Science. Previously he was a post-doc
with Mark Steedman lab in the University of Edinburgh. Omri earned his
PhD from the Hebrew University, where he co-developed the UCCA scheme
along with his doctoral advisor Ari Rappoport. Omri's research focuses
on semantics, including semantic representation and parsing, as well
as corpus annotation and semantic approaches to text-to-text
generation. See http://www.cs.huji.ac.il/~oabend/ for details.