Weekly Seminar: Marianna J. Martindale
February 06, 2025
Details
Where: TMCB 1170
When: February 13th, 2025
Speaker: Marianna J. Martindale
Talk Title
Risks and Mitigations for Machine Translation Enabled Triage in Intelligence Analysis
Abstract
The perception of high Machine Translation (MT) quality, based largely on performance translating between high resource languages, has led to its adoption for a variety of use cases. However, MT is still imperfect and can even be misleading in its errors. Government use cases such as those involving law enforcement or intelligence are uniquely risky because government action or inaction based on incorrect information can cause harm to individuals or national security. One such use case is MT for triage in intelligence analysis. In this talk, I will first introduce the MT-enabled triage use case and some of the risks in the context of intelligence analysis as well as U.S. intelligence community policies implicated by this use case that may partially mitigate these risks. I will then present a user study with intelligence analysts to measure their baseline performance on this type of MT-enabled triage task and the effectiveness of providing additional translation outputs as a practical intervention to further mitigate risk.
Biography
Marianna J. Martindale received her PhD in Information Studies from the University of Maryland in August 2024 after previously receiving her BS in Computer science from BYU (2003) and MS in Linguistics (Computational) from Georgetown University (2010). Since 2003, she has worked for the US Government supporting, building, and deploying machine translation systems. She is currently serving as the secretary on the board of the Association for Machine Translation in the Americas. Her research interests include user-centered machine translation evaluation and how to help users calibrate their trust in generative AI model outputs.