Truth in Modeling: How Close Do We Actually Get to What We Want to Model?

Abstract: To model any construct (e.g., hate speech, stance, logic, mental health, etc.), we need to make a few assumptions: the construct should exist, it should be modelable, and we should be able to reliably detect it. In a word, it should be true. However, modeling truth in practice is much less of a binary notion than we would want it to be. In this talk about completed and ongoing work in my lab, I will show how many of our assumptions about construct validity and truth need to be reexamined for annotation, benchmark data sets, LLM output, LLM rankings, and even publication. I will show evidence of assumptions gone wrong, and methods on how to control for and mitigate that effect. Importantly, all that should not discourage us from seeking to model the truth. But it should make us even more keenly aware of the subtle (and not so subtle) effects that make this noble pursuit even more difficult than it already was. Speaker Bio: To model any construct (e.g., hate speech, stance, logic, mental health, etc.), we need to make a few assumptions: the construct should exist, it should be modelable, and we should be able to reliably detect it. In a word, it should be true. However, modeling truth in practice is much less of a binary notion than we would want it to be. In this talk about completed and ongoing work in my lab, I will show how many of our assumptions about construct validity and truth need to be reexamined for annotation, benchmark data sets, LLM output, LLM rankings, and even publication. I will show evidence of assumptions gone wrong, and methods on how to control for and mitigate that effect. Importantly, all that should not discourage us from seeking to model the truth. But it should make us even more keenly aware of the subtle (and not so subtle) effects that make this noble pursuit even more difficult than it already was.