Title: Learning and Reasoning with Semantic Vectors

Abstract: Learning distributional similarities from text is well-established,
and has been used in many tasks including information retrieval, word sense
disambiguation, lexical acquisition and translation. Different mathematical
techniques are available, including singular value decomposition, latent Dirichlet
analysis, and random projection, with different semantic and computational
properties that make them appropriate for different tasks. However, for many years
these methods could be used only to detect symmetric "similarity" relationships:
due to the lack of syntactic or directed relationships, they have often been
described as "bag-of-words?" methods.

In the last few years, research has taken a big step further, in learning
relationships beyond similarity. Linguistically, this is stimulated by taking
word-order, or deeper syntactic and semantic relationships, into account.
Mathematically, it is implemented by using more sophisticated operators, such as
permutation or convolutions of coordinates, to represent language composition.

In this talk, we will briefly survey this area, and discuss a specific application
of
semantic vectors to inference in the medical domain. We use semantic vectors to
learn
representations for different objects and their relationships, and test this model
by
applying it to an analogical reasoning test, of the form "A is to X, as B is to
what?"

Finding Schizophrenia's Prozac: Emergent Relational Similarity in Predication
Space
Cohen. T. Widdows, D. Schvaneveldt, RW. Rindflesch, T. Proceedings of the Fifth
International Symposium on Quantum Interaction, Aberdeen, UK, 2011.
http://www.puttypeg.net/papers/schizophrenias-prozac.pdf

According to results to date, these methods are both robust and computationally
efficient. The combination of learning and reasoning within the same model is a
challenge that applies to many disciplines, and is perhaps a key ingredient in
building intelligent technology.