We share collected rationale annotations with socio-demographic information, and code, so others will hopefully continue this line of work. It will be in github very soon 😊

The work indicates that there can be model biases that do not surface in measures of fairness wrt standard task performance, and might only be revealed by considering rationales.

In sentiment classification, we find that model rationales tend to align better with the rationales given by older individuals.

In our study we collect annotations and assess whose rationales a model, or explain. method aligns with, and find that even when people agree on the label, and the model is equally right across diverse annotator groups, it may not be equally right in its reasoning:

Biases and spurious correlations are typically detected by the use of explainability methods, but these are themselves evaluated against relativity subjective human rationales 🤔

Are you interested in #NLProc explainability and fairness? Then you’re in luck - the preprint for our #ACL2023 paper is out! https://t.co/GNtbadstmd Work done with the lovely @LauraCabelloPi and Anders Søgaard. Summary 🧵

Happy to share that our paper "Being Right for Whose Right Reasons?"🧐has been accepted to #ACL2023NLP! More details coming soon, stay tuned! Work with @Ternethorn and Anders Søgaard

This, ladies and gentlemen, is what those long pesky reviwerer sign-up forms can do for you. Really proud of what we were able to do at ACL23, and hope that area chairs will find it useful for meta-reviews. If any other organizers are interested, we can share the forms and code.

I am defending my PhD dissertation soon! Come along if you are interested in NLP biases, crowdsourcing, explainability and peer-review 🤸‍♀️ A zoom link will also become available at some point 🎥

The paper:

Tomorrow I have much to celebrate, I’m presenting our paper on NLP conference peer-review at #NAACL2022 (session 5A, 14:15 PDT) & it’s my birthday! 🥳👵🏻

I wrote a paper with @annargrs! And she made a very nice thread about it:

Hey #NLProc, looking for some Friday afternoon reading? Check the link below for a collection of blog posts about #EMNLP2021 papers, spanning topics from bias, robustness, summarization, low-resource, and so much more! 👇 Thanks to all the authors 🥳 https://t.co/LuzLMCrBj0

📢 Dear #NLProc authors: Have your papers mostly been sent to the right reviewers? If not - what kinds of papers? What kind of problems ensued? Share your experience by Dec 7: https://t.co/FHMt02t9pU 🔁 RT appreciated! /1

Today I presented my very first paper at @_starsem! The paper is about Spurious Correlations in Cross-Topic Argument Mining. Read: https://t.co/TomKTrpCUb or watch the presentation: https://t.co/3tFr2U7gCL OR watch the 3.48 min (!) easy digest version: https://t.co/YIwzO8xWbq :)

How are our daily lives and practices shaped by digital data, and how is this evident during the Corona-crisis? Take a look at @sdbreslin's piece in @IndependentNL: https://t.co/QowCHagner #SocialDataScience

Hand washing problem solved in Texas

Join us and Bloxhub on September 23rd at Københavns Universitet - University of Copenhagen for a timely and critical talk with Roger McNamee (@Moonalice) - early Facebook investor, former mentor to Mark Zuckerberg and musician. Register here: https://t.co/sO8tRiVuEZ (1/2)

When at a conference in Belgium, of course there is beer #MMSym

Hansen, Chair of @DKNudgeNet shares some thoughts on nudging relative to traditional regulation measures