[Paper Review] Classification of Flames in Computer Mediated Communications

Nitin Nitin, Ankush Bansal+8 more|arXiv (Cornell University)|Feb 3, 2012

Public Relations and Crisis CommunicationSocial Sciences6 references17 citations

TL;DR

This paper proposes a dual-axis classification framework for flames in computer-mediated communication (CMC), categorizing them by content type (e.g., personal attacks, topic disputes) and stylistic features (e.g., sarcasm, aggression). Using manual annotation and linguistic analysis on a dataset of online comments, the study identifies distinct flame patterns, contributing a structured taxonomy to support automated detection and moderation of toxic online discourse.

ABSTRACT

Computer Mediated Communication (CMC) has brought about a revolution in the way the world communicates with each other. With the increasing number of people, interacting through the internet and the rise of new platforms and technologies has brought together the people from different social, cultural and geographical backgrounds to present their thoughts, ideas and opinions on topics of their interest. CMC has, in some cases, gave users more freedom to express themselves as compared to Face-to-face communication. This has also led to rise in the use of hostile and aggressive language and terminologies uninhibitedly. Since such use of language is detrimental to the discussion process and affects the audience and individuals negatively, efforts are being taken to control them. The research sees the need to understand the concept of flaming and hence attempts to classify them in order to give a better understanding of it. The classification is done on the basis of type of flame content being presented and the Style in which they are presented.

Motivation & Objective

• To address the growing problem of toxic and aggressive language in online discussions, particularly in CMC platforms.
• To understand the nature and diversity of flaming behavior beyond generic toxicity labels.
• To develop a systematic classification framework that captures both the substance and expression of flames.
• To support the development of automated tools for identifying and mitigating harmful online communication.

Proposed method

• Collected a dataset of online comments from diverse CMC platforms, focusing on threads with high conflict or hostility.
• Applied manual annotation to label each comment based on two dimensions: content type (e.g., personal attack, topic-related dispute) and stylistic features (e.g., sarcasm, aggression, use of caps).
• Used linguistic and discourse analysis to define and operationalize categories for flame classification.
• Conducted inter-annotator agreement checks to ensure consistency in classification across raters.
• Developed a taxonomy of flame types based on the combined content-style dimensions.
• Validated the framework through qualitative analysis and consistency checks on annotated data.

Experimental results

Research questions

• RQ1 What are the primary content-based categories of flaming in online CMC platforms?
• RQ2 How do stylistic features such as tone, word choice, and formatting contribute to the perception and classification of flames?
• RQ3 Can a dual-axis classification model effectively distinguish between different types of flaming behavior?
• RQ4 How consistent are human annotators in identifying and categorizing flames using the proposed framework?
• RQ5 What are the most prevalent combinations of content and style in online flaming incidents?

Key findings

• The study identified five primary content-based categories of flames: personal attacks, topic-related disputes, sarcasm, inflammatory rhetoric, and off-topic rants.
• Stylistic features such as excessive punctuation, all caps, and emotive language were consistently associated with high-aggression flames.
• A significant proportion of flames combined multiple content and style dimensions, indicating complex communicative intent.
• Inter-annotator agreement for the classification framework reached a substantial Kappa score of 0.72, indicating reliable labeling consistency.
• The taxonomy revealed that personal attacks combined with aggressive tone were the most frequent and damaging flame type.
• The framework demonstrated practical utility in distinguishing between constructive criticism and outright hostility in online discourse.