Laura Aina

Hi Bangkok 🇹🇭, if you are attending @aclmeeting stop by our poster "Factual Confidence of LLMs: on Reliability and Robustness of Current Estimators". Our amazing @AmazonScience intern Matéo Mahaut is in-person in 🇹🇭: 14:00–15:30 Session 4, Aug 12 (today) #ACL2024NLP #NLProc

Our AWS science team in sunny Barcelona is looking for an intern. Great opportunity to do research and learn about exciting topics related to large language models and generative AI! amazon.jobs/en/jobs/222231…

Our results point to efficient alternatives to the usage of large LMs and labeling a large amount of training examples to build effective NLP applications. In particular, prompting + KD allows to keep both data and compute cost (of inference) low. 3/3

We study how performance varies across model and train set sizes, when adapting pre-trained language models for classification tasks through different training procedures (fine-tuning vs. prompting; w/ vs. w/o knowledge distillation) 2/3

Upcoming AACL 2022 paper "Performance-Efficiency Trade-Offs in Adapting Language Models to Text Classification Tasks" with @nickvosk and @roiblanco arxiv.org/abs/2210.12022 @AmazonScience 1/3

On the 11th of April, 3pm-5pm, join us in our mini-workshop on Referential Information in Deep Learning Models 🤖, with Raquel Fernàndez and Anna Rogers. More details at this link: upf.edu/web/amore/home…… @annargrs @raquel_dmg @colt_upf @TraduccioUPF @ERC_Research

✨🥹

NEW: On the 25th of March, 3pm-5pm, our mini-workshop on linguistic ambiguity will be hybrid. Join us at this zoom link: google.com/url?q=https://… More details about the event: upf.edu/web/amore/home…… @AllysonEttinger @colt_upf @TraduccioUPF @ERC_Research

Looking forward to this amazing workshop about linguistic ambiguity! (and to defending my PhD thesis on that topic earlier that day🤞)

Finally, I’m looking for a new PhD candidate to join the group! Due to funding constraints, this position must start in 2021 (Master’s degree required). Only a few days left to apply: tinyurl.com/Phd-dmg-2021 There may be open postdoc positions next year, stay tuned 😎

Our results suggest that predictability affects mention type (pronoun/name/description) and length: more informative context-->less informative expression. But a joint role with other discourse factors (e.g., recency, syntactic position) helps to explain reference production 4/4

We use computational estimates of referent predictability to scale the analysis on more data. We modify a coreference resolution architecture to give us predictions on "masked" mentions. e.g.: "Laura congratulated Xixian because [MASK] got a paper accepted." 3/4

I'm happy to share "Does referent predictability affect the choice of referential form? A computational approach using masked coreference resolution", joint work with @xixian18, Gemma Boleda @amoreupf and @MatthijsWestera, accepted at #CoNLL '21 arxiv.org/abs/2109.13105 1/4

To whoever needs to read this: When you realize your goals are beyond what you can handle, it's okay to take a step back and slow the pace. And if you need to put your plans fully on hold to recover, be patient and try not to put too much pressure on yourself. ❤️ 6/6

For various reasons, during the course of this project I've had to prioritize my health on multiple occasions, leading to frequently readjust plans and expectations. It wasn't an easy process, which makes me even happier that I can now share this work. 5/6

I'm happy to announce that my paper with @tallinzen "The Language Model Understood the Prompt was Ambiguous: Probing Syntactic Uncertainty Through Generation" has been accepted at #BlackboxNLP '21 arxiv.org/abs/2109.07848 1/6