Understanding the Causes and Consequences of Online Hate Speech

• In Event: Hate Speech: Causes and Reactions

Abstract

This study examines the causes and consequences of online hate speech against women by systematically analyzing comments on one of the largest online news platforms in the world. Our study creates a novel, online news participation-tracking dataset from January 2011 to October 2020 using NAVER News, the largest news-aggregating service in South Korea.
Our goals are threefold. First, we provide the first-ever descriptive network of comments on the Naver. To our knowledge, this research constitutes the largest network of comments on any news site to date. The data identifies the number of commenting users, the frequency of comments, the level of interaction among commenters (as indicated by subcomments on the same post), and the degree of homogeneity among news sources that receive similar attention by commenters.
Second, we analyze the content of comments on Naver. Here we use Naver’s current method for flagging hate speech, “Cleanbot,” and employ machine-learning to replicate their algorithm. This enables us to see how Naver’s algorithm itself has changed (if at all) and, more importantly, to apply a similar algorithm to the eight years of comments on Naver before Cleanbot was implemented to identify hate speech. Using these data we compare the networks of communication among those who engage in hate speech. We also use the natural experiment of Naver instituting Cleanbot to see how flagging inflammatory speech causally impacts hate speech and news consumption habits more generally.
Last, we examine the relationship between hate speech on Naver and hate crimes in Korea. The field of communication is rich with studies examining the relationship between news coverage and public opinion. But this literature rarely examines the link between coverage and action. Moreover, very few, isolated studies can examine the link between news content and behavior at the individual level. We simply do not know what news specific individuals read, watch, or hear. (The most notable exception is the Surveys of Consumers question, “Have you heard anything about the economy in the past month?” which focuses exclusively on the economy and does not ask the respondent to identify the news source.) The Naver dataset is uniquely suited for these questions because it identifies not only the stories that are most read by all individuals in Korea, but precisely the stories that people read and then comment on. That is, we can know what news stories each commenter actually read right before they engaged in hate speech. Moreover, we can ask: does hate speech lead or lag (or both) hate crimes? Hate crimes against women have become increasingly common in Korea over the past decade, and it is an important context to better understand these connections. While data of this scope is unavailable in the United States (because people do not gather their news from a single online news aggregator on any scale like they do in Korea), we believe many of the questions and findings will have broad applications outside of the Korean context.

Authors

• Georgia Kernell, UCLA

• Seonhye Noh, UCLA