Detecting Online Echo Chambers Gradients with Bidirectional Encoder Representations from Transformers

社交媒体和网络世界提供了一个与人联系的工具，可以被认为是我们作为一个物种所做的最大的事业之一，在地球仪的任何地方与其他人在真实的时间联系。通过这种扩展的对话以及为后代存档和存储这些对话的能力，这些平台在分享比以往任何时候都更多的信息方面提供了社会和政治利益。然而，尽管越来越多的人能够在网上表达自己，西方世界在政治上却进一步分裂。无论是在政治还是非政治时代，影响选民思想的力量都是所有政党的宝贵财富。无论这是有意还是无意，大数据和推荐算法的时代已经改变了平台影响观众的方式，代价是平衡观点和思想的多样性，导致攻击性和部落/反社会行为。本文并不是要防止或减少观点，而是要识别缺乏常规外部意见的群体，而是吸收加强自己的观点，导致对对立群体或个人的毫无根据的疏远和概括。为了缓解这个问题，首先需要确定回声室在网络内的位置。我们提出了一种新的方法在检测在线回声室应用计算机视觉边缘检测算法，以确定符合群体的大小和位置。这些组将基于用户的节点交互，通过使用来自变压器的双向编码器表示（BERT）模型来预测交互因素，例如对话的主题，用户对主题的立场，用户的情绪和他们的曝光。然后，这些模型的值将绘制出一个用户网络，我们将使用人工智能通过地面实况评估其检测回声室的准确性。

Social Media and the online world has provided a tool to connect with people and could be considered one of the largest undertakings we have taken as a species, to connect with others in real time, anywhere around the globe. With this expanded conversation and the ability to archive and store these conversations for generations to come, these platforms have provided benefits socially and politically in sharing more information than ever before. However, whilst more people are able to express themselves online, the western world has further become politically divided. During both political and non-political times, the power of influencing voter thoughts is a valuable asset for all political parties that can attain it. Whether this is or is not intentional, the age of big data and recommendation algorithms has changed the way that platforms influence viewers at the cost of balanced viewpoints and the diversity of thought, leading to aggressive and tribal/antisocial behaviour. This paper is not proposing to prevent or lessen viewpoints, but instead to identify groups that lack regular external opinions and instead assimilate views that reinforce their own, causing unfounded alienation and generalisations towards an opposing group or individual. In an attempt to mitigate this problem, it is first necessary to determine where echo chambers are within a network. We propose a novel approach in detecting online echo chambers by applying computer vision edge detection algorithms to determine the size and location of conforming groups. These groups will be based on user's node interactions by using Bidirectional Encoder Representations from Transformers (BERT) models to predict interaction factors such as the topic of conversation, the user's stance on the topic, the user's sentiment and their exposure. These models' values will then map out a network of users where we will assess its accuracy in detecting an echo chamber via ground truth using Artificial Intelligence.