simplified overview of our aligned probing setup, where we join the behavioral and internal evaluation of LMs' toxicity
LMs that "know more" about toxicity are less toxic!
Our #TACL 📄 connects behavior and internals:
💠 LMs amplify toxicity beyond humans
💠 Information about toxicity peaks in lower layers
💠 Bypassing these layers increases toxicity
More details👇 #NLProc #interpretability (1/🧵)
Automatic Speech Transcription Boosts Speaker Attribution Accuracy
ASR transcripts can match human transcripts for speaker attribution, with accuracy dropping only marginally as word‑error rates rise. The paper, accepted to TACL, posted 11 July 2025. getnews.me/automatic-speech-transcr... #speechrecognition #tacl
🎉 Congratulations again to the recipients of the TACL Best Paper Award for 2024 🏆and the TACL Test of Time Award for 10 years 🏆!
Your outstanding contributions exemplify excellence and lasting impact in computational linguistics.
#TACL #ACL2025NLP #TestOfTime
OOD evaluation pipeline of self-rationalization, and OOD datasets categories considered in the paper.
Can LLMs generate explanations for datasets without such annotations? 🧠
We tested model explanations across 19 datasets (NLI, fact-checking, hallucination detection) to see how well they self-rationalize on completely unseen data.
#LLMs #Explainability #ACL2025 #TACL
View from the Graben in Vienna with UKP Lab Logo and #ACL2025
🎉 The countdown to #ACL2025 in Vienna is on - and we’re excited to contribute 10 accepted papers and two #TACL articles from the UKP Lab!
In the days leading up to the conference, we’ll be sharing each of our contributions, including links to the arXiv preprints, code repositories, and more.
(1/🧵)
🚨 Two Papers @aclmeeting.bsky.social! 🚨
🎉 1×ACL Main + 1×TACL
🎥 What Is That Talk About? A Video-to-Text Summarization Dataset for Scientific Presentations (ACL)
🧠 Explanatory Summarization with Discourse-Driven Planning (TACL)
Thanks to my amazing co-authors! 🙏
#LLMs #NLProc #TACL #ACL #ACL2025
Prompting LMs is not enough to quantify their linguistic competence! Meet the Holmes🔎 benchmark at #EMNLP2024 #TACL or 👉🧵
💠Meta-study of current literature
💠Comprehensive coverage of LMs and phenomena
💠Analysis of LM size, architecture, and instruction tuning
holmes-benchmark.github.io
