Radicalization and Polarisation in New Media: The Case of YouTube

拥有比任何其他社交网站更多的用户（Auxier和Anderson， 2021）， YouTube是现代社会中最具影响力的新媒体形式之一。作为一个平台，它是一个娱乐的游乐场和一个可以交换思想和发展新观点的集市。然而，YouTube作为集市的角色引起了很多争议。有人认为，该平台可能会导致观点的激进化和两极分化。它也被指责为假新闻的温床。不幸的是，几乎没有经验证据来证实这些说法。错误信息与平台上的激进化或两极分化之间的相互作用尚未得到彻底探索。该项目使用新颖的文本和视觉数据，以及机器学习（ML）和自然语言处理（NLP）的最新进展来正式评估这些说法。使用Ribeiro等人（2019）中分配给YouTube频道的类别作为标签，我将使用ML算法来预测每个视频的激进化程度。我将增强缩略图，以提高预测能力。使用视频内容作为变量在这一领域是一个新颖的引入，而之前的研究主要集中在使用评论。将视频内容添加到这种分析中可以改进以前使用的研究激进化的方法。该项目的第一部分涉及激进化。YouTube的推荐系统是用户在该平台上浏览视频的一个关键方面。坊间证据表明，YouTube的算法往往会将观众引向更极端的内容（Tufekci, 2018）。尽管如此，很少有研究调查该算法对用户接触更极端内容的影响。这项研究将从识别多个主题（例如，疫苗接种，全球变暖，选举）的种子视频开始，并设计一个算法来导航由YouTube算法提供的推荐网络。缩略图图像和评论将从每个视频收集。从这些文本和可视化数据出发，我们将构建激进化的措施。然后，这些将被用来正式量化我们在网络中移动时激进化的变化。该项目的第二部分涉及两极分化。使用文本数据，我们试图构建YouTube上的两极分化措施。这将使我们能够确定两极分化是如何随着时间的推移而演变的，无论是在现有渠道内还是通过创建新渠道。对广泛和密集边缘变化的分析使我能够描绘出两极分化是如何随着时间而演变的。为了衡量两极分化，我们采用了类似于根茨科等人（2019）使用的方法，他们使用美国国会演讲来衡量政治两极分化。为了捕捉视频创作者对内容的选择，我们引入了一个选择模型。记录文本数据用于度量视频内容。然后通过估计该模型的选择概率来构建极化度量。这可以解释为观察者可以轻易地猜测视频的意识形态起源，因为创作者选择了单一的短语。我们概括了genzkow等人的衡量标准，允许它纳入多维的意识形态起源，这更适合YouTube。早期的两极分化措施被认为是有偏见的（Gentzkow et al., 2019）。这是因为选择集相对于观察到的短语的选择更大。为了解决这个问题，我们将使用诸如省略和正则化估计器之类的策略。

Having more users than any other social networking site (Auxier and Anderson, 2021), YouTube is one of the most influential forms of new media in modern society. As a platform it is a playground for entertainment and an agora in which ideas can be exchanged and new opinions can be developed. However, YouTube has caused much controversy with its role as an agora. It has been suggested that the platform can contribute to radicalization and polarisation of opinions. It is also blamed to be a hotbed of fake news. Unfortunately, there is little empirical evidence to substantiate these claims. The interaction between misinformation and radicalisation or polarisation on the platform hasn't been thoroughly explored.This project uses novel text and visual data, along with recent advances in machine learning (ML) and natural language processing (NLP) to formally assess these claims.Using the categories assigned to YouTube channels in Ribeiro et al. (2019) as labels, I will use ML algorithms to predict radicalisation measures of each video. I will augment thumbnail images to increase predictive power. The use of video content as a variable is a novel introduction in this area, whereas previous studies have mainly focused on using comments. Adding video content to this analysis improves upon previously used methods to study radicalization. The first part of the project concerns radicalisation. YouTube's recommendation system is a key aspect of the way in which users explore videos on the platform. Anecdotal evidence has suggested that YouTube's algorithm tends to lead viewers to content that is more extreme (Tufekci, 2018). In spite of this, few studies have investigated the effects of the algorithm on users' exposure to more extreme content.This research will begin by identifying seed videos across a number of topics (e.g., vaccination, global warming, elections) and designing an algorithm to navigate through the recommendation network provided by the YouTube algorithm. Thumbnail images and comments will be collected from each video. From this text and visual data, measures of radicalization will be constructed. These will then be used to formally quantify how radicalization changes as we move through the network. The second part of the project concerns polarisation. Using text data, we seek to construct measures of polarisation on YouTube. This will allow us to identify how polarisation has evolved over time, both within existing channels and through the creation of new channels. The analysis on changes across the extensive and intensive margins enable me to paint a picture of how polarisation has evolved over time.In order to measure polarisation we follow a method similar to that used in Gentzkow et al. (2019), who used US congressional speech to measure political polarisation. We introduce a choice model in order to capture content choices made by video creators. Transcript text data is used to measure video content. Polarisation measures are then constructed by estimating choice probabilities from this model. These can be interpreted as the ease with which an observer can guess the ideological origin of a video, given the creator's single choice of phrase. We generalise Gentzkow et al.'s measure, allowing it to incorporate multidimensional ideological origins, which are more suited to YouTube. Earlier polarisation measures are argued to be biased (Gentzkow et al., 2019). This is because the choice set is large relative to the choice of phrases that is observed. In order to address this, we will use strategies such as leave-out and regularised estimators.