Pride, Love, and Twitter Rants: Combining Machine Learning and Qualitative Techniques to Understand What Our Tweets Reveal about Race in the US
Pride, Love, and Twitter Rants: Combining Machine Learning and Qualitative Techniques to Understand What Our Tweets Reveal about Race in the US
复制标题
DOI：
10.3390/ijerph16101766
复制
发表时间：
2019-05-02
影响因子：
--
通讯作者：
Nguyen, Quynh C.
中科院分区：
文献类型：
--
作者：
Nguyen, Thu T.;Criss, Shaniece;Nguyen, Quynh C.
目的：描述使用种族相关术语的推文情绪的变化，并确定表征与种族相关的社会气候的主题。研究方法：我们应用随机梯度下降分类器对2015-2016年使用种族相关术语的1，249，653条美国推文进行了情感分析。为了评估准确性，手动标签与计算机标签进行了比较，随机子集的6600条推文。我们对随机抽取的2100条推文进行了定性内容分析。结果：计算机标签与手工标签的一致率为74%。与提及亚洲人（17.7%）和西班牙裔（17.5%）的推文相比，提及中东群体（12.5%）或黑人（13.8%）的推文的积极情绪最低。定性内容分析显示，大多数推文由以下类别表示：消极情绪（45%），积极情绪，如文化自豪感（25%）和导航关系（15%）。虽然所有推文都使用一个或多个与种族有关的术语，但非贬损性或中心主题与种族无关的负面情绪推文很常见。结论：这项研究利用相对未开发的社交媒体数据来开发一种新的区域层面的社会背景（情感得分）衡量标准，并强调了这项工作中的一些挑战。衡量社会环境的新方法可能会加强对社会背景和健康的研究。
Objective: Describe variation in sentiment of tweets using race-related terms and identify themes characterizing the social climate related to race. Methods: We applied a Stochastic Gradient Descent Classifier to conduct sentiment analysis of 1,249,653 US tweets using race-related terms from 2015-2016. To evaluate accuracy, manual labels were compared against computer labels for a random subset of 6600 tweets. We conducted qualitative content analysis on a random sample of 2100 tweets. Results: Agreement between computer labels and manual labels was 74%. Tweets referencing Middle Eastern groups (12.5%) or Blacks (13.8%) had the lowest positive sentiment compared to tweets referencing Asians (17.7%) and Hispanics (17.5%). Qualitative content analysis revealed most tweets were represented by the categories: negative sentiment (45%), positive sentiment such as pride in culture (25%), and navigating relationships (15%). While all tweets use one or more race-related terms, negative sentiment tweets which were not derogatory or whose central topic was not about race were common. Conclusion: This study harnesses relatively untapped social media data to develop a novel area-level measure of social context (sentiment scores) and highlights some of the challenges in doing this work. New approaches to measuring the social environment may enhance research on social context and health.