fdata-03-00003_The Datafication of Hate: Expectations and Challenges in Automated Hate Speech Monitoring.xml

Hate speech has been identified as a pressing problem in society and several automated approaches have been designed to detect and prevent it. This paper reports and reflects upon an action research setting consisting of multi-organizational collaboration conducted during Finnish municipal elections in 2017, wherein a technical infrastructure was designed to automatically monitor candidates' social media updates for hate speech. The setting allowed us to engage in a 2-fold investigation. First, the collaboration offered a unique view for exploring how hate speech emerges as a technical problem. The project developed an adequately well-working algorithmic solution using supervised machine learning. We tested the performance of various feature extraction and machine learning methods and ended up using a combination of Bag-of-Words feature extraction with Support-Vector Machines. However, an automated approach required heavy simplification, such as using rudimentary scales for classifying hate speech and a reliance on word-based approaches, while in reality hate speech is a linguistic and social phenomenon with various tones and forms. Second, the action-research-oriented setting allowed us to observe affective responses, such as the hopes, dreams, and fears related to machine learning technology. Based on participatory observations, project artifacts and documents, interviews with project participants, and online reactions to the detection project, we identified participants' aspirations for effective automation as well as the level of neutrality and objectivity introduced by an algorithmic system. However, the participants expressed more critical views toward the system after the monitoring process. Our findings highlight how the powerful expectations related to technology can easily end up dominating a project dealing with a contested, topical social issue. We conclude by discussing the problematic aspects of datafying hate and suggesting some practical implications for hate speech recognition.

仇恨言论已成为社会亟待应对的严峻问题，现有诸多自动化技术方案已被研发用于检测并遏制此类言论。
本文针对2017年芬兰市政选举期间开展的多组织协作行动研究场景进行报告与反思：该场景中搭建了一套技术基础设施，用于自动监测候选人的社交媒体动态中的仇恨言论。
该研究场景为我们开展双重维度的调查提供了充足条件。
其一，本次多组织协作为探究仇恨言论如何被建构为技术问题提供了独特视角。本项目采用监督机器学习（supervised machine learning）技术开发了一套性能达标且可行的算法解决方案：我们测试了多种特征提取与机器学习方法，最终选用词袋特征提取（Bag-of-Words）与支持向量机（Support-Vector Machines）的组合方案。
但自动化检测方案往往需要进行大幅简化，例如采用简易分类标尺对仇恨言论进行分级，且依赖基于词汇的分析方法；而现实中，仇恨言论是兼具语言与社会属性的复杂现象，存在多样的语气与表现形式。
其二，本次行动研究场景让我们得以观察与机器学习技术相关的各类情感反馈，包括参与者对该技术的期许、畅想与担忧。基于参与式观察、项目产出物与文档、项目参与者访谈以及公众对该检测项目的线上反馈，我们梳理出参与者对高效自动化方案的诉求，以及算法系统所宣称的中立性与客观性程度。
但在监测流程结束后，参与者对该系统的批评声音有所增多。我们的研究结果揭示出：当项目涉及极具争议性的社会热点议题时，围绕技术的过高期望极易主导整个项目的走向。
最后，本文讨论了将仇恨言论数据化所存在的问题，并为仇恨言论识别任务提出若干实践启示。