Effect of circular features on model training for IR Example

author:

date: 30. April 2021

download: https://github.com/StatNLP/empirical_methods


F1 on training data

F1 on test data

BASELINE vs EXTENDED Model (students: all features)

Baseline

Extended Model

Nullification

TEACHER: extended model, STUDENT: no citation

Search for Circularity

Baseline system (without circular features)

Top 10

Selected model

BIEF System (with circular features)

Top 10

Selcted model

Ablation study 1

Remove citation information (set all values to zero)

Ablation study 2

Remove all non-circular information (set all values to zero)

Teacher models

Baseline

BIEF model (has access to all features)

BIE Model (patial circular teacher)

BI and BIE Model (all patial circular teachers)