NLP Highlights · Allen Institute for Artificial Intelligence

142 - Science Of Science, with Kyle Lo

·49 min·3 clips
Kyle connects science of science to "reducing income inequality" and says concentrated research effort shapes which problems get studied.
1. NLP Highlights episode 142 centers on Kyle Lowe and the science of science topic. 2. Bridger and Pradeep Dazeghi host Kyle Lowe, who works at the Allen Institute for Artificial Intelligence and studies science of science. 3. The episode asks what science of science is for and how it can improve the process of science itself. 4. Kyle defines science of science as computational sociology that uses computational or logical methods to study how scientific communities share ideas, publish, and build on each other. 5. He says the field also studies individual scientists, technology availability, and funding, which makes the area "super broad." 6. Kyle says communities often study their own fields, including NLP, chemistry, and biology, rather than one unified science-of-science community. 7. He names James Evans and Karl Bergstrom as examples of researchers who study science across multiple fields. 8. Kyle traces the roots of the area back to Kuhn in the 1960s and says people were asking what it means to do science even in the 1930s. 9. He says his own motivation is to build useful tools for scientists, but only after understanding what is useful for a field and not just for an individual. 10. Kyle points to the Nature 2019 paper by Lingfei Wu, Dashun Wang, and James Evans, "Large Teams Develop and Small Teams Disrupt Science and Technology," as a testable kind of science-of-science result. 11. He says that work could inform an NSF grant maker or a university administrator deciding how to move money toward smaller groups. 12. Kyle describes science of science as a way to understand how ideas are transmitted and how scientists adopt information. 13. He says that understanding can affect how he writes, designs experiments, and chooses problems to work on. 14. He argues that field-specific work can be more actionable, and he gives an arXiv example with Lucy Wang and Hanchen Cao about hyperlink use in preprints. 15. That project looks at how scientific work increasingly lives in linked artifacts like data, code, and videos that are not preserved like papers. 16. Kyle says the same project also asks how often old hyperlinks are already gone when researchers try to revisit them. 17. The conversation shifts to NLP as an enabler for science of science, especially through data access, PDF cleaning, citation graphs, and large-scale information extraction. 18. Kyle says open data and efficient models matter because a loop that goes from data annotation to model building to large-scale inference is still slow and expensive. 19. The discussion stays technical but conversational, with both hosts pressing for concrete examples and Kyle answering in a reflective, practical style. 20. Researchers in NLP, ML, or open-science infrastructure would get the most from this episode, while people wanting a light nontechnical listen might skip it.

As heard by us

An attentive look at using NLP to make science easier to study.

The episode treats science of science as a practical problem: how computational methods can help study the process of science, and where language tools can make that work easier to handle.

Read the full review in PlayNext →

Why you'd press play

Press play for a grounded tour of how NLP can help study science itself.

Read the full recommendation in PlayNext →
Listen to the show on