[Paper Review] When a Tweet is Actually Sexist. A more Comprehensive Classification of Different Online Harassment Categories and The Challenges in NLP

Sima Sharifirad, Stan Matwin|arXiv (Cornell University)|Feb 27, 2019

Hate Speech and Cyberbullying DetectionComputer Science12 references20 citations

TL;DR

This paper proposes a refined, multi-category classification of online harassment on Twitter, introducing five distinct types: indirect harassment, information threat, sexual harassment, physical harassment, and not sexist. Using supervised machine learning on a manually annotated dataset, it demonstrates that a more nuanced taxonomy improves the detection of subtle and complex forms of sexism beyond traditional hostile/benevolent dichotomies, offering a foundation for more effective NLP-based hate speech detection systems.

ABSTRACT

Sexism is very common in social media and makes the boundaries of freedom tighter for feminist and female users. There is still no comprehensive classification of sexism attracting natural language processing techniques. Categorizing sexism in social media in the categories of hostile or benevolent sexism are so general that simply ignores the other types of sexism happening in these media. This paper proposes a more comprehensive and in-depth categories of online harassment in social media e.g. twitter into the following categories, "Indirect harassment", "Information threat", "sexual harassment", "Physical harassment" and "Not sexist" and address the challenge of labeling them along with presenting the classification result of the categories. It is preliminary work applying machine learning to learn the concept of sexism and distinguishes itself by looking at more precise categories of sexism in social media.

Motivation & Objective

• To address the limitations of existing binary classifications (hostile vs. benevolent sexism) in capturing the full spectrum of online harassment.
• To develop a more comprehensive and granular framework for categorizing online harassment on social media, particularly on Twitter.
• To investigate the challenges in labeling and detecting diverse forms of sexism using NLP techniques.
• To provide a foundation for improved machine learning models that can detect subtle and indirect forms of sexist content.

Proposed method

• The authors define five distinct categories of online harassment: indirect harassment, information threat, sexual harassment, physical harassment, and not sexist.
• A dataset of Twitter posts was manually annotated by human raters using the proposed taxonomy to ensure labeling consistency and accuracy.
• Supervised machine learning models were trained on the annotated dataset to classify tweets into the five predefined categories.
• The study evaluates model performance using standard NLP metrics such as precision, recall, and F1-score to assess classification effectiveness.
• The authors address labeling challenges by analyzing inter-annotator agreement and discussing ambiguities in classifying subtle or indirect sexist language.
• The approach emphasizes context-aware classification, recognizing that sexist content often appears in indirect or coded forms.

Experimental results

Research questions

• RQ1 How can online harassment on Twitter be classified into more granular and meaningful categories beyond the traditional hostile/benevolent dichotomy?
• RQ2 What are the key challenges in labeling subtle and indirect forms of sexism using NLP techniques?
• RQ3 How effective are machine learning models in distinguishing between different types of online harassment, including sexual and physical threats?
• RQ4 To what extent does a more comprehensive taxonomy improve the detection of sexist content compared to binary classification?
• RQ5 What are the main sources of ambiguity and disagreement in annotating online harassment, and how can they be mitigated?

Key findings

• The proposed five-category taxonomy enables more precise identification of various forms of online harassment, including indirect and threat-based behaviors often missed by binary models.
• The study demonstrates that indirect harassment and information threats are prevalent but under-recognized forms of online sexism.
• Inter-annotator agreement was moderate, highlighting the complexity and subjectivity involved in labeling subtle or context-dependent sexist language.
• Machine learning models achieved measurable performance on the new taxonomy, showing that nuanced classification is feasible with proper annotation.
• The results suggest that current NLP systems may significantly under-detect sexist content due to reliance on oversimplified classification schemes.
• The paper establishes that a more comprehensive framework is essential for developing effective, context-aware hate speech detection tools in social media.