Computer Studies Bldg. Room 209
One reason text understanding is hard is because only a fraction of the
knowledge the author intended to convey is explicitly stated. As
a result, for computers to correctly interpret text and "fill in the gaps",
a vast amount of lexical and world knowledge is needed. This challenge is
present both in the small (Does a sentence reasonably follow from a text?
The task of recognizing textual entailment), and in the large (What is a
coherent representation of multiple texts about the same topic?). In this
talk I will present some of our forays at Boeing into this task, in particular
the Recognizing Texual Entailment challenge, in which we attempt to leverage
three vast sources of knowledge (WordNet, 12 million paraphrases, and
55 million "tuples"). I will discuss what worked, what didn't, and
tentatively suggest some further small steps forward toward the AI dream
of machines that can read for themselves.
Bio: Peter Clark is an Associate Technical Fellow in Boeing's Network Systems
Technology Organization in Seattle, WA, where he leads projects in the areas
of knowledge-based systems, commonsense reasoning, and natural language
processing, including for Vulcan's Project Halo and previously for IARPA's
AQUAINT Program.
Refreshment will be provided at 10:45AM