Inference with Linear Constraints
for Natural Language Tasks

View of Solving NLP Problems

Weaknesses of Pipeline Model

Global Inference with Classifiers

Ideal Inference

Inference with Classifiers

Setting (1/2)

Setting (2/2)

Outline

Problem Setting

Integer Linear Programming

Phrase Identification – A simple case

LP Formulation – Linear Cost

LP Formulation – Linear Constraints

LP Formulation

Outline

Entity/Relation Recognition

Problem Setting

LP Formulation – Indicator Variables

LP Formulation – Cost Function

LP Formulation – Linear Constraints

LP Formulation

Experiments – Data

Experimental Results – F₁

Decision-time Constraint

Outline

Semantic Role Labeling

Example

Argument Types

Examples

Algorithm

I. Find Potential Arguments

II. Arguments Type Likelihood

Details – Phrase-level Classifier

What is a Good Assignment?

Inference

LP Formulation – Linear Cost

Linear Constraints (1/2)

Linear Constraints (2/2)

Results on Perfect Boundaries

Results

Lesson

Outline

Generality

Computational Issues (1/2)

Computational Issues (2/2)

Current and Future Work (1/3)

Current and Future Work (2/3)

Current and Future Work (3/3)

Summary

References

Thank you