On Cross-Dataset Generalization in Automatic Detection of Online Abuse

Isar Nejadgholi, Svetlana Kiritchenko


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.
Anthology ID:
2020.alw-1.20
Volume:
Proceedings of the Fourth Workshop on Online Abuse and Harms
Month:
November
Year:
2020
Address:
Online
Editors:
Seyi Akiwowo, Bertie Vidgen, Vinodkumar Prabhakaran, Zeerak Waseem
Venue:
ALW
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
173–183
Language:
URL:
https://aclanthology.org/2020.alw-1.20
DOI:
10.18653/v1/2020.alw-1.20
Bibkey:
Cite (ACL):
Isar Nejadgholi and Svetlana Kiritchenko. 2020. On Cross-Dataset Generalization in Automatic Detection of Online Abuse. In Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 173–183, Online. Association for Computational Linguistics.
Cite (Informal):
On Cross-Dataset Generalization in Automatic Detection of Online Abuse (Nejadgholi & Kiritchenko, ALW 2020)
Copy Citation:
PDF:
https://aclanthology.org/2020.alw-1.20.pdf
Optional supplementary material:
 2020.alw-1.20.OptionalSupplementaryMaterial.zip
Video:
 https://slideslive.com/38939537