Presentation #212.04 in the session Public Engagement and Education Research.
Two challenges of teaching science to non-science majors are training them to write coherently about science and giving them the ability to distinguish between legitimate science and misinformation online. We are using a machine learning approach to tackle both issues. Neural networks have succeeded in a public benchmark task to extract facts and corresponding verification evidence from a large set of Wikipedia articles. Our application is to train neural networks to identify claim-evidence pairs in undergraduate science writing. First, undergraduate science majors select online articles on topics such as climate change and evolution, with equal numbers that are legitimate science and misinformation. They each classify all of the articles as either “real” or “fake” and then manually identify claims and evidence to support those claims in each article. This information is used to train the neural network. The trained neural network is applied to writing assignments in large introductory science classes for non-science majors, with the goal of training students in how to recognize legitimate science arguments. The tool will be used by instructors for formative assessment and to give students constructive feedback on their writing. Initial results suggest trained neural networks can recognize scientific misinformation with better than 90% reliability. The eventual goal is a browser extension that will judge the legitimacy of any web page about science on the Internet.