Natural language research announcements

[Rajhans Samdani <rsamdan2 AT illinois.edu>]

Hi all, 

Just a gentle reminder regarding today's talk by Amir Globerson 
(http://www.cs.huji.ac.il/~gamir/) from Hebrew University. The talk is in 
3405 SC at 2 
pm. Following are the title and abstract of his talk.

Title: Learning with Approximate inference - From LP Relaxations to 
Pseudo-Max 
Approaches 

Abstract: 
Supervised learning problems often involve the prediction of complex 
structure labels, 
such as sequences (e.g., POS tagging) or trees (e.g., dependency parsing). To 
achieve high accuracy in these tasks, one is often interested in introducing 
complex 
dependencies between label parts. However, this can result in prediction 
problems 
that are NP hard. A natural approach in these cases is to use tractable 
approximations of the prediction problem. 
In this talk I will present our recent work on using approximate inference 
for structured 
prediction tasks. I will describe linear programming (LP) relaxations for the 
problem, 
and show highly scalable algorithms for learning using these relaxations. I 
will next 
introduce a simpler approach, called ‘psuedo-max’  learning, and show that it 
is 
consistent for separable problems under certain conditions, and has empirical 
performance that is similar to LP relaxations. 
I will conclude by addressing the problem of finding the K best solutions in 
such 
problems, and show a new class of relaxations that has theoretical guarantees 
and 
works well in practice. 

Base on joint work with: Ofer Meshi, David Sontag, Menachem Fromer and Tommi 
Jaakkola 


Hoping to see you there. 
Best, 
Rajhans 


Rajhans Samdani,
Graduate Student,
Dept. of Computer Science,
University of Illinois at Urbana-Champaign.