Natural language research announcements

["Horn, Eric Bailey" <erichorn AT illinois.edu>]

When: Tuesday, July 24 @ 4pm

 

Where: 3403 Siebel Center

 

Speaker: Elisabetta Fersini of the University of Milano-Bicocca

 

Title: Soft Constrained Inference for Conditional Random Fields

 

Abstract:

The discovering of semantic information embedded within natural language documents can be viewed as a decision making process aimed at assigning a sequence of labels to a set of interdependent variables. This problem can be modeled through a stochastic process involving both hidden variables (labels) and observed variables (textual cues). One of the most recent and promising stochastic processes for dealing with text labeling is represented by Conditional Random Fields (CRFs). In this talk their potential for Named Entity Recognition tasks will be presented, by focusing our attention on the introduction – during the inference phase - of extra knowledge learned directly from training data. In particular, by exploiting a recent Integer Linear Programming formulation introduced by Roth and Yih, a two stages approach for dealing with the automatically extracted extra-knowledge will be presented. The proposed approach allows some constraints, which model logic relationships that should be satisfied during the output prediction, to be violated.

Experimental results on both real and benchmark data show the potentiality of learning constraints from data as well the proposed integer linear programming constrained inference.

 

Bio:

Dr. Elisabetta Fersini is a postdoctoral research fellow at the Department of Informatics, Systems and Communication at the University of Milano-Bicocca. She is involved in several research projects both national initiatives and funded by the European Community.

The research activity is mainly focused on:

1) Machine Learning: algorithms and models for supervised and unsupervised learning, with particular interest to Support Vector Machines, Bayesian Networks and Partitioning Clustering.

2) Relational Learning: structural uncertainty modeling, algorithms and models for supervised and unsupervised learning in relational domains, with particular interest to Relational Clustering, Probabilistic Relational Models and Conditional Random Fields.

The research activity has found application in various domains, including:

-              World Wide Web: document clustering, document classification, information extraction

-              Life Sciences: prediction of anticancer drug responses

-              e-Justice: integrated decision support systems for in and out of court proceedings

 

Please email (joanlund AT illinois.edu) your availability if you are interested in meeting with this speaker.