Developing robust methods for the measurement of online social media populations

2016年6月，令人不安的证据浮出水面，显示谷歌的搜索引擎是“种族主义者”。争论的焦点是搜索词“三个白色青少年”和“三个黑人青少年”返回的截然不同的图像。 前一次搜索返回了青少年运动和闲逛的健康照片。 后者归还了嫌疑犯的照片。 谷歌为搜索结果进行了辩护，表示他们的搜索算法是根据网页内容及其访问和链接的频率来确定的。 谷歌--这家运行着数百万人每天在互联网上查找信息的软件的公司--声称他们搜索结果中的任何此类偏见“从根本上说是一个社会问题"。这就是本质上的难题。 我们期待谷歌搜索-沿着从Yelp到OkCupid到Facebook的无数其他在线服务-来反映世界。 当机器学习算法成为传播不受欢迎的刻板印象和边缘化少数观点的工具时--我们称之为“算法偏见”--我们就有问题了。解决算法偏见提出了两个挑战。 首先，我们必须发现偏见何时存在。 其次，如果存在偏见，并认为有必要纠正，我们需要调整算法的性能的方法。 这些是我们在本提案中考虑的核心任务。*我们会分三个阶段处理这些问题。 首先，我们将开发方法来帮助检测算法偏见。 第二，与领域专家合作，我们将使用我们的检测方法对检测仇恨言论和识别政治倾向的算法中存在的偏见进行详细描述。 利用这些发现，我们将接近本提案的最后一个主题，即纠正算法偏见的方法-通过修改算法本身或调整算法训练的数据。随着越来越多的服务转移到网上，我们必须确保这些系统反映社会的愿望，而不是其最坏和最掠夺性的元素。我们希望这项工作将为公司提供可行的解决方案，以确保无偏见的系统，并希望我们的研究结果能够促进这一重要主题的进一步工作。

In June 2016, troubling evidence surfaced showing Google's search engine to be “racist.” At issue were the radically different images returned by the search terms “three white teenagers” and “three black teenagers.” The former search returned wholesome pictures of teenagers playing sports and hanging out. The latter returned mugshots. Google defended the search results, indicating that their search algorithms were informed by web content and the frequency with which it was accessed and linked. Google - the company that runs software used daily by millions of people to find information on the internet - asserted that any such bias in their search results were “fundamentally a societal problem.”***And there, in essence, is the conundrum. We look to Google search – along with myriad other online services from Yelp to OkCupid to Facebook – to reflect the world. And in doing so, we give it the power to reinforce the social structures in it. When machine learning algorithms become vehicles for propagating undesirable stereotypes and marginalizing minority views – which we call “algorithmic prejudice” - we have a problem.***Addressing algorithmic prejudice presents two challenges. First, we must detect when a prejudice exists. Second, where a prejudice exists and is deemed necessary to correct, we require methods for adjusting the algorithm's performance. These are the central tasks we consider in this proposal.***We will tackle these problems in three phases. First, we will develop methods to aid in the detection of algorithmic prejudice. Second, in collaboration with domain experts, we will use our detection methods to do a detailed characterization of the prejudice present in algorithms for detecting hate speech and identifying political orientation. Using these findings, we will approach the final topic of this proposal, which is methods for correcting algorithmic prejudice – either by modifying the algorithms themselves or by adjusting the data that the algorithms are trained on.***As increasing numbers of services migrate online, we must ensure that these systems reflect the aspirations of society, not its worst and most predatory elements. Our hope is that this work will provide viable solutions for companies to ensure prejudice-free systems and that our findings stimulate further work on this important topics.