Learning Mental Representations from the Web

Current knowledge extraction methods are often targeted at a particular
set of relations or named entities, typically using textual pattern
matching. A more comprehensive approach to knowledge extraction would
ideally have the same broad coverage, generative power and task-agnostic
applicability that is seen in human intelligence. For this to be
realized, we need word representations that can be deconstructed into
semantic components (e.g. to make inferences), which can be
assembled into larger phrasal meanings, and which correspond to human
behaviour. I will present web-derived meaning representations that have
some of these properties, and evaluate them by comparing their
predictions with several kinds of human judgements, and recordings of
brain activity.


Brian Murphy is a scientist at the Department of Machine Learning,
Carnegie Mellon University, working on computational semantics and the
neuroscience of language.
He completed an M.Phil and PhD in Computational Linguistics at Trinity
College Dublin, followed by a postdoc at the Centre for Mind/Brain
Sciences of the University of Trento.