←
Katie Keith261 posts
Katie Keith avatar

Katie Keith

@katakeith

NLP and CSS researcher. Assistant Professor at @WilliamsCollege. @allen_ai alum, @UMass_NLP alum. she/her

◷ Joined September 2014▧ 34 Photos and videos
Katie Keith avatar
Katie Keith@katakeith·20 Aug 2024

These were some of the many reasons I chose Williams. No institution is perfect, but this one checks a lot of boxes. Hope you can join us! (4/4)

Katie Keith avatar
Katie Keith@katakeith·20 Aug 2024

3. 5 out of 13 tenure-track CS faculty are women 4. Student body: 52% receive need-based financial aid (all grants) and over 40% identify as US students of color. (3/4)

Katie Keith avatar
Katie Keith@katakeith·21 Jun 2024

At @nlpandcss, Naoki Egami “Using Large Language Model Annotations for the Social Sciences" talk, lots of questions about the assumptions of "annotators as ground truth." I think this is a really important question! Often disagreement is an inherently useful signal.

Post attachment
Katie Keith avatar
Katie Keith@katakeith·19 Jun 2024

I'm a bit biased 😀, but I found this a great read for those interested in text+causal inference!

This tweet is unavailable

Katie Keith avatar
Katie Keith@katakeith·19 May 2024

Exciting set of ideas and people coming from American Causal Inference Conference (ACIC) in Seattle this past week! Smaller conferences = more ideas sparked. Fantastic watching @r0ntu and my @WilliamsCollege undergraduate students share their work!

Post attachmentPost attachmentPost attachmentPost attachment
Katie Keith avatar
Katie Keith@katakeith·15 May 2024

We also implement an example evaluation pipeline and walk through finite data considerations with a novel, real-world RCT from AI2’s @SemanticScholar platform. This dataset consists of approximately 70k observations with text data as high-dimensional covariates.

Katie Keith avatar
Katie Keith@katakeith·15 May 2024

In this work, we subsample RCTs to create confounded observational datasets while using the average causal effects from the RCTs as ground-truth. We contribute a new sampling algorithm “RCT rejection sampling” and provide theoretical guarantees that causal identification holds.