Extractive Adversarial Networks: High-Recall Explanations for Identifying Personal Attacks in Social Media Posts

TL;DR

This paper presents an adversarial approach to generate high-recall extractive explanations for neural text classifiers, specifically targeting the detection of personal attacks in social media comments, with emphasis on model bias and interpretability.

Contribution

It introduces an adversarial layer to enhance extractive explanations and highlights the importance of setting a semantic default behavior for improved detection of personal attacks.

Findings

Adversarial layer improves explanation recall.

Explicit bias manipulation enhances attack detection.

Validation with human annotations confirms effectiveness.

Abstract

We introduce an adversarial method for producing high-recall explanations of neural text classifier decisions. Building on an existing architecture for extractive explanations via hard attention, we add an adversarial layer which scans the residual of the attention for remaining predictive signal. Motivated by the important domain of detecting personal attacks in social media comments, we additionally demonstrate the importance of manually setting a semantically appropriate `default' behavior for the model by explicitly manipulating its bias term. We develop a validation set of human-annotated personal attacks to evaluate the impact of these changes.