Comment-level Topic Drift Analysis in the Reddit Corpus

This paper investigates how online discourse evolves by measuring whether topic discussions shift meaningfully through a high-dimensional semantic embedding space over time, distinguishing genuine conceptual drift from random stochastic noise.

The authors analyze 12.7 billion Reddit comments spanning from 2006 to 2022 using a three-stage pipeline. First, they map short texts into a 384-dimensional vector space using a pretrained transformer model. Second, they partition the data into monthly time windows and apply k-means clustering to identify topic clusters. Third, they align these clusters across time using dimensionality reduction and hierarchical clustering to trace topic trajectories, evaluating their movement against a random-walk null model.

The application reveals that while recreational topics like music and sports remain stationary in embedding space, socially and politically contentious themes undergo substantial directional drift. Furthermore, inter-topic distances shift systematically, demonstrating that concepts once treated as semantically distinct can converge or diverge over a multi-year horizon.

Understanding semantic drift at scale provides researchers and moderators with quantitative tools to measure cultural shifts, track the evolution of online polarization, and improve downstream tasks such as content moderation and discourse forecasting.