Artificial Intelligence

Structured Learning With Inexact Inference

Friday, January 27, 2012, 11:00am - 12:00pm PDTiCal
11th Floor Conf. Rm (#1135)
Liang Huang


Structured learning is an important subfield of machine learning with wide applications in natural language processing, bioinformatics, computer vision and related fields. In most of these applications, exact inference is often intractable, for example in parsing and machine translation. Therefore learning with inexact inference is a fundamental problem.

This work develops a general theory of structured perceptron learning under inexact inference. We propose variants of the structured perceptron algorithm under a general “violation-fixing” framework that guarantees convergence. This framework subsumes previous remedies including “early update” as special cases, and also explains why standard perceptron may fail with inexact search. We also propose new update methods within this framework which learn better models with dramatically reduced training times on state-of-the-art part-of-speech tagging and incremental parsing systems. As a by-product, we achieved the best part-of-speech tagging accuracy to date.

Home Page:

« Return to Events