Tracking the Diffusion of Named Entities

Leon Derczynski, Matthew RowearXiv 2017 · arXiv

Introduction

Information diffusion studies have traditionally focused on "deterministic" signals—static strings like URLs, hashtags, or group memberships. While these are computationally easy to track, they represent only a fraction of human discourse. The most meaningful components of conversation are named entities (e.g., "Conor McGregor" or "Hungary"). However, tracking entities is notoriously difficult due to the noisy nature of social media text, characterized by misspellings, abbreviations, and informal slang. Standard Named Entity Recognition (NER) tools often fail in these environments.

This paper addresses three primary challenges: accurately extracting entities from noisy Reddit data, characterizing the "shape" of entity spread, and predicting future adoptions.

Methodology

NER for Social Media

To handle the scale and noise of Reddit, the researchers developed a multi-stage pipeline. They utilized a Conditional Random Fields (CRF) model enhanced with Brown clustering. A significant technical innovation involved avoiding "shearing"—the practice of shortening bitstrings representing word clusters. By using the full path to the root cluster, the model preserved more hierarchical information, improving its ability to generalize across informal text.

Cascade Analysis

The researchers modeled conversation trees where entities were mentioned. Using graph isomorphism (specifically the Cordella algorithm), they identified and categorized the various shapes these cascades take. This allowed for a structural comparison between entity diffusion on Reddit and other platforms like the blogosphere.

Figure 4: Top-20 cascade shapes and their rank distribution, illustrating the "deep and narrow" nature of Reddit conversations.

General Threshold Model (GTM)

To predict adoption, the authors adapted the GTM framework into a parallelized version using Apache Spark and HBase. They tested three "influence constructs":

  1. Entity Propagation: Direct exposure to a neighbor mentioning the entity.
  2. Interactions: The frequency of direct replies or interactions with a neighbor.
  3. Community Homophily: Shared participation in the same subreddits.

Datasets

The study utilized a massive 140GB dataset containing approximately 1.7 billion Reddit posts from the site's inception to July 2015. For NER training, they used a blend of the CoNLL 2003 (newswire) corpus and Twitter datasets (Ritter and W-NUT 2015). To evaluate performance, the authors manually annotated a 3,708-token "gold-standard" Reddit development set.

Key Insights and Results

NER Performance and Cluster Efficacy

The tuned NER system achieved an $F_2$ score of 66.84 on Reddit text, significantly outperforming the Stanford NER baseline of 47.81. The researchers prioritized $F_2$ (recall) because missing an entity is more detrimental to diffusion studies than occasional false positives.

A key finding was that "in-genre" clusters are superior. Using a smaller amount of Reddit-specific data to build word clusters was more effective than using much larger datasets from newswire or other social media platforms.

Figure 2: F2 performance across different cluster data sources, highlighting the impact of data volume and genre on NER accuracy.

Cascade Dynamics: Deep and Narrow

While information in the blogosphere often spreads "wide and shallow" (many sources linking to one), Reddit entity cascades are "deep and narrow." They typically manifest as long, thin chains of conversation with minimal branching.

Furthermore, the researchers observed a "Zero-Exposure" Phenomenon. Unlike hashtags, which often require multiple exposures before adoption, the most common number of prior exposures for a Reddit entity adoption was zero. This suggests that entities often emerge spontaneously within a community or are brought in from external sources.

Interaction Trumps Propagation

The strongest predictor of adoption was not simple exposure (Entity Propagation) but the intensity of prior interactions with neighbors. Reddit is a discussion-driven platform; the "stickiness" of a relationship is the primary driver of micro-topic flow. The interaction-based model achieved a Micro-ROC value of 0.755.

Temporal Dynamics

The average "influence window" (the time taken for an entity to propagate between users) was approximately 1.2 years. This highlights the long-term "stickiness" and slow-burn nature of Reddit community discourse.

Conclusion

This research demonstrates that tracking named entities provides a more nuanced view of information diffusion than simple hashtags. By prioritizing interaction intensity and in-genre data for NER, researchers can better understand how complex ideas move through digital communities. Future work will scale this analysis to thousands of entities and explore multivariate diffusion to see how the adoption of related entities correlates over time.

Summary updated

Abstract

Existing studies of how information diffuses across social networks have thus far concentrated on analysing and recovering the spread of deterministic innovations such as URLs, hashtags, and group membership. However investigating how mentions of real-world entities appear and spread has yet to be explored, largely due to the computationally intractable nature of performing large-scale entity extraction. In this paper we present, to the best of our knowledge, one of the first pieces of work to closely examine the diffusion of named entities on social media, using Reddit as our case study platform. We first investigate how named entities can be accurately recognised and extracted from discussion posts. We then use these extracted entities to study the patterns of entity cascades and how the probability of a user adopting an entity (i.e. mentioning it) is associated with exposures to the entity. We put these pieces together by presenting a parallelised diffusion model that can forecast the probability of entity adoption, finding that the influence of adoption between users can be characterised by their prior interactions -- as opposed to whether the users propagated entity-adoptions beforehand. Our findings have important implications for researchers studying influence and language, and for community analysts who wish to understand entity-level influence dynamics.
BibTeX
@misc{derczynski2017tracking,
  author = {Leon Derczynski and Matthew Rowe},
  title = {{Tracking the Diffusion of Named Entities}},
  year = {2017},
  eprint = {1712.08349},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1712.08349},
}
Figure