Nick McKenna

What do you mean? PhD in #NLProc #semantics @EdinburghNLP, formerly @BrownCSDept

[Nick McKenna @lexicalizer]

@emnlpmeeting @jgwinnup Hi, this blog post is still a bit confusing and overly harsh sounding. Is it meant to read like this? "if ANY author has at least 3 *CL pubs, then AT LEAST ONE author must review AT LEAST 3 papers, or else desk-reject; exception applies if NO authors have AT LEAST 3 papers"

[Asa Cooper Stickland @AsaCoopStick]

It was fun watching GPT-4 fail miserably on this dataset! A truly herculean effort by @idavidrein and co means the questions are super high quality and interesting. I hope to see many scalable oversight evaluations on this soon...

[Nick McKenna @lexicalizer]

@rsltgy1 @gozde_gul_sahin @nikita_moghe @EdinburghNLP Hi @gozde_gul_sahin I have some availability after tomorrow if you still need someone? I can DM you if you follow back since I'm not verified :)

[Nikita Moghe @nikita_moghe]

🚨 New paper alert! 🚨
We all know that LLMs can be integrated with tools and they can be personalised through clever prompts or more recently "custom instructions"?
But what happens when we evaluate these instructions more systematically?
https://t.co/AyrT3mW3h9
(1/N)

[Nick McKenna @lexicalizer]

@LukasSchaefer96 @EdinburghNLP @aaclmeeting Thank you ☺️

[Nick McKenna @lexicalizer]

We are so grateful to win the Best Paper Award at #AACL2023 🥳😄

Our work “Smoothing Entailment Graphs with Language Models” bridges symbolic and explainable NLI into the era of language models.

Thank you to my coauthors, @EdinburghNLP , and @aaclmeeting !

[AACL 2026 @aaclmeeting]

🎉 Best Paper! 🎉
The 'Smoothing Entailment Graphs with Language Models' paper by Nick McKenna, Tianyi Li, Mark Johnson, and Mark Steedman wins the Best Paper Award at #AACL2023!
🎉 Big congrats! 🎉
https://t.co/W86IeexBuM

[Nick McKenna @lexicalizer]

@tttthomasssss @EdinburghNLP @aaclmeeting Thank you!!!

[Nick McKenna @lexicalizer]

@rsltgy1 @EdinburghNLP @aaclmeeting Of course! You always know best 😂

[Nick McKenna @lexicalizer]

@jinheonbaek @EdinburghNLP @aaclmeeting Thanks Jinheon! :)

[Nick McKenna @lexicalizer]

@mjavadhosseini @EdinburghNLP @aaclmeeting Thank you so much!

[Nick McKenna @lexicalizer]

In beautiful #Bali 😍

[AACL 2026 @aaclmeeting]

🎉 Best Paper! 🎉
The 'Smoothing Entailment Graphs with Language Models' paper by Nick McKenna, Tianyi Li, Mark Johnson, and Mark Steedman wins the Best Paper Award at #AACL2023!
🎉 Big congrats! 🎉
https://t.co/W86IeexBuM

[Tianyi Li @TianyiLi_Teddy7]

Check out this exciting joint work with @lexicalizer and folks, to find the two prominent factors causing LLMs to hallucinate on NLI tasks! 👀

[Nick McKenna @lexicalizer]

These models fail to generalize in ways a human would. We offer our experimental controls as tools to evaluate future models for their ability to generalize in NLI tasks beyond artifacts seen in training.

[Nick McKenna @lexicalizer]

Better late than never:

We (@TianyiLi_Teddy7 and I) are excited to announce our paper will appear in #EMNLP23 findings: "Sources of Hallucination by Large Language Models on Inference Tasks"

Preprint link: https://t.co/AbyjjNgqqj

@emnlpmeeting @InfAtEd @EdinburghNLP

[Nick McKenna @lexicalizer]

LLM performance appears very good when sample labels conform to these biases, but performance drops drastically when samples do not conform. We show that these models use such biases to approximate language inference, but these biases have no relation to language meaning.

[Nick McKenna @lexicalizer]

And of course, thank you @EdinburghNLP ! #AACL2023

[Nick McKenna @lexicalizer]

Very pleased to announce that our paper, "Smoothing Entailment Graphs with Language Models" has been accepted to #AACL2023 🎉

Many thanks to co-authors @TianyiLi_Teddy7, Mark Johnson, and Mark Steedman

Preprint link: https://t.co/ZkeBeCPRcn

[Nick McKenna @lexicalizer]

Our paper explores this question with a symbolic model for NLI, an Entailment Graph.

An off-the-shelf language model is used to approximate statements missing from the EG using known statements, completing a transitive chain. This improves both recall, and even precision.