On Cross-Dataset Generalization in Automatic Detection of Online Abuse

TL;DR

This paper investigates the challenges of cross-dataset generalization in online abuse detection, highlighting biases in datasets and proposing methods to improve robustness and transferability of models.

Contribution

It identifies topic and task biases affecting generalization, and demonstrates that removing biased topics enhances cross-dataset performance without harming in-domain accuracy.

Findings

Benign examples in Wikipedia Detox are biased towards platform-specific topics.

Removing biased topics improves cross-dataset generalization.

Unsupervised topic modeling helps identify and reduce dataset biases.

Abstract

NLP research has attained high performances in abusive language detection as a supervised classification task. While in research settings, training and test datasets are usually obtained from similar data samples, in practice systems are often applied on data that are different from the training set in topic and class distributions. Also, the ambiguity in class definitions inherited in this task aggravates the discrepancies between source and target datasets. We explore the topic bias and the task formulation bias in cross-dataset generalization. We show that the benign examples in the Wikipedia Detox dataset are biased towards platform-specific topics. We identify these examples using unsupervised topic modeling and manual inspection of topics' keywords. Removing these topics increases cross-dataset generalization, without reducing in-domain classification performance. For a robust dataset design, we suggest applying inexpensive unsupervised methods to inspect the collected data and downsize the non-generalizable content before manually annotating for class labels.