For well over a century, the journal Science Education has been publishing articles about the teaching and learning of science. These articles represent more than just a repository of past work: they have the potential to offer insights into both the history of science education as well as well as the dynamics of field-specific change. It can be difficult, however, for educators, researchers, reformers, and policymakers to grasp the nuances of over 100 years of scholarship given the overwhelming amount of textual material. To address this problem, we have used latent Dirichlet allocation, an automated machine-learning algorithm from the field of natural language processing, to perform an automated literature review and classification of the corpus of work in Science Education. Using this technique, we have classified research in the journal into 21 distinct topics, falling into three thematic groups: science content topics, teaching-focused topics, and student-focused topics. We have also quantified the rise and fall of these topics and groups over time, and used them to begin to extract insight into the development of the field, including the effects of national policy changes on topics of interest to the research community, the interrelationships between different research topics, and the effects of intellectual cross-pollination. Based on this analysis, we argue that this technique shows great promise for even larger-scale analyses of educational literature and other textual data.

中文翻译：

---

在过去的 100 年里，科学教育发生了怎样的变化？使用自然语言处理的分析

一个多世纪以来，《科学教育》杂志一直发表有关科学教学和学习的文章。这些文章不仅仅代表过去工作的资料库：它们有可能提供对科学教育历史以及特定领域变化动态的见解。然而，鉴于大量的文本材料，教育工作者、研究人员、改革者和政策制定者可能很难掌握 100 多年学术研究的细微差别。为了解决这个问题，我们使用了潜在狄利克雷分配（一种来自自然语言处理领域的自动机器学习算法）来对科学教育工作语料库进行自动文献审查和分类. 使用这种技术，我们将期刊中的研究分为 21 个不同的主题，分为三个主题组：科学内容主题、教学主题和学生主题。我们还量化了这些主题和群体随着时间的推移的兴衰，并利用它们开始深入了解该领域的发展，包括国家政策变化对研究界感兴趣的主题的影响，之间的相互关系不同的研究课题，以及智力异花授粉的影响。基于这种分析，我们认为这种技术对于对教育文献和其他文本数据进行更大规模的分析显示出巨大的希望。