All Times ET
Program is Subject to Change
Coding Text Survey Responses Using Machine Learning: 5 Years In (307971)*Alexander Measure, Bureau of Labor Statistics
Keywords: machine learning, natural language processing, text classification, autocoding
Five years ago Bureau of Labor Statistics staff read and manually classified hundreds of thousands of written descriptions of work related injury and illness each year. Today, more than two thirds of these codes are assigned by a deep neural network, which evaluations suggest is more accurate than trained humans. In this presentation we discuss how we addressed some of the many challenges inherent in this transition including how we built these machine learning systems, how we evaluate their performance, how we decide when and how to use them, and how we monitor and maintain them to continually improve performance.