Tutorial 10: Estimating causal effects of aspects of language with noisy proxies скачать в хорошем качестве

Tutorial 10: Estimating causal effects of aspects of language with noisy proxies

Description: Does the politeness of an email or a complaint affect how quickly someone responds to it? This question requires a causal inference: how quickly would someone have responded to an email had it not been polite? With observational data, causal inference requires ruling out all the other reasons why polite emails might be correlated with fast responses. To complicate matters, aspects of language such as politeness are not labeled in observed datasets. Instead, we typically use lexicons or trained classifiers to predict these properties for each text, creating a (probably noisy) proxy of the linguistic aspect of interest. In this talk, I’ll first review the challenges of causal inference from observational data. Then, I’ll use the motivating example of politeness and response times to highlight the specific challenges to causal inference introduced by working with text and noisy proxies. Next, I’ll introduce recent results that establish assumptions and a methodology under which valid causal inference is possible. Finally, I’ll demonstrate this methodology: we’ll use semi-synthetic data and adapt a text representation method to recover causal effect estimates. Tutorial host: Dhanya Sridhar Slides: https://nlp-css-201-tutorials.github.... Code: https://colab.research.google.com/dri... This is part of a larger tutorial series, NLP+CSS 201: Beyond the basics, which is organized by Ian Stewart and Katherine Keith. Website: https://nlp-css-201-tutorials.github....