[Paper Review] Abusive Language Detection with Graph Convolutional Networks

Pushkar Mishra, Marco Del Tredici+2 more|arXiv (Cornell University)|Apr 5, 2019

Hate Speech and Cyberbullying Detection18 references46 citations

TL;DR

The paper introduces a heterogeneous graph approach and graph convolutional networks to model online communities and linguistic behavior for abusive language detection, achieving superior results over baselines.

ABSTRACT

Abuse on the Internet represents a significant societal problem of our time. Previous research on automated abusive language detection in Twitter has shown that community-based profiling of users is a promising technique for this task. However, existing approaches only capture shallow properties of online communities by modeling follower-following relationships. In contrast, working with graph convolutional networks (GCNs), we present the first approach that captures not only the structure of online communities but also the linguistic behavior of the users within them. We show that such a heterogeneous graph-structured modeling of communities significantly advances the current state of the art in abusive language detection.

Motivation & Objective

• Motivate the need to move beyond purely structural community features by incorporating linguistic behavior of authors.
• Propose a two-graph representation (community graph and extended author-tweet graph) to capture both social structure and text content.
• Develop a semi-supervised GCN-based method to learn author profiles that reflect linguistic behavior and community structure.
• Compare GCN-based methods against strong baselines using a subset of the Waseem & Hovy Twitter dataset.
• Demonstrate with experiments that GCN-based author profiling outperforms prior methods.

Proposed method

• Represent the community as two graphs: a homogeneous author graph and an extended heterogeneous graph with author and tweet nodes.
• Use node2vec on the homogeneous graph as a baseline to obtain author profiles.
• Apply a two-layer Graph Convolutional Network on the extended graph to learn embeddings for all nodes, using labeled tweet nodes for semi-supervised training.
• Input features are binary bag-of-words vectors; authors’ features aggregate over all their tweets.
• Train with cross-entropy loss on labeled tweet nodes and extract 200-d embeddings from the first GCN layer for authors.
• Compare with baselines: logistic regression on character n-grams (lr), lr with author profiles from node2vec on both graphs (lr+auth, lr+extd), and lr+gcn.

Experimental results

Research questions

• RQ1 Can a heterogeneous graph that includes both authors and their tweets improve abusive language detection beyond author-only graphs?
• RQ2 Does incorporating linguistic signals via a GCN improve author profiling and tweet classification compared to node2vec baselines?
• RQ3 How do GCN-based profiles perform across racism, sexism, and overall classification metrics?

Key findings

• The lr+gcn method significantly outperforms all other methods (p<0.05, paired t-test).
• Author profiles from the GCN capture both community structure and linguistic behavior, addressing misclassifications that vanilla author profiles miss.
• GCN alone yields high recall for sexism but lower precision, and struggles with racism due to a small number of abusive authors contributing mixed content.
• Combining lr with GCN (lr+gcn) mitigates racism issues by incorporating tweet-level character n-gram features.
• Node2vec-based author profiles on the extended graph offer little to no gain over the community graph alone, unlike the GCN approach.
• Overall, lr+gcn achieves the best macro-averaged precision, recall, and F1 across racism, sexism, and clean classes.