Title: Utilizing recursive Matrix-Vector Spaces for TAC Knowledge Base Population

Title: Utilizing recursive Matrix
-
Vector Spaces for TAC Knowledge Base
Population


Goal:
Improve the LSV TAC KBP system by deriving features from recursive neural
network matrix vector spaces


Background:
Relation extraction systems try to extract informat
ion like “Obama was born
in Honolulu” from text. The last step in this is often a classifier deciding whether the
relation is present or not in a sentence. The features of such a classifier can use properties
that are trained unsupervised on a large backgr
ound corpus like automatically trained word
classes.


Description:
The neural network model suggested by Socher et al. provides a very
convincing meaning representation. The vector
-
matrix representation has properties of
compositionality that go beyond sim
ple vector combination suggested by Mitchell and
Lapata. The plan is to use this richer semantic representation in a state of the art TAC
Knowledge
-
Base
-
Population system.


Tasks:

1.

Understand the Socher et. al paper

2.

Explore the software made available be So
cher

3.

Train the recurrent neural network on a news wire background corpus

4.

Extract features from the neural network

5.

Build a relation classifier to the TAC KBP task

6.

Explore alternative vector
-
matrix representations and ways to combine them

7.

Explore alternative

grammar formalisms (e.g. a simple regular grammar)


References:



Socher, Richard and Huval, Brody and Manning, Christopher D. and Ng, Andrew Y.,
Semantic Compositionality Through Recursive Matrix
-
Vector Spaces, EMNLP 2012



Jeff Mitchell and Mirella Lapata.
2010.
Composition in Distributional Models of
Semantics
. To appear in
Cognitive Science
.



LSV
-
Tac paper (Not yet on web page)