🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Ahhh I miss these two

Legendary camera man #49Unbeaten

Our supercool paper got accepted at #EMNLP2020 ! See y'all in (a Zoom meeting with a scenic picture of) Punta Cana! #NLProc

An interesting argument by Ev (like in 'evidence', not Eve) Fedorenko 🇺🇦 in her #EMNLP2021 keynote: humans use language for communication, not for thought—and so language models should not be judged for their ability to reason. I would add: we should also not use lang.models as proxies for cognitive abilities.

Every course needs its guest rockstar... This year we have Alfredo Canziani! ✌️

What a pleasure it's been to teach 3 hours of self-supervised energy-based models at Sapienza Università di Roma with a super warm audience, split between physically present and remotely connected students. Thanks, Simone Scardapane, for hosting me today. 😋😋😋

New work! Humans appear to learn similarly for different modalities and so should machines! data2vec uses the same self-supervised algorithm to train models for vision, speech, and nlp. Paper: ai.facebook.com/research/data2… Blog: ai.facebook.com/blog/the-first… Code: github.com/pytorch/fairse…

Save the date!

I love the idea of 4-day workweeks. Think about the amount of research that can be done on 3 day weekends.

progression of prompt engineering research

New work on autoregressive language models for retrieval! We train our model, SEAL (Search Engines with Autoregressive LMs) to produce text snippets occurring somewhere in the corpus, in relevant documents. We use the generated ngrams to rank documents in the corpus. Thread 👇

Differentiable Search Indices showed that autoregressive transformers can operate as “search engines” to some extent, but not for full scale corpora. FM-Indices change this! Really enjoyed this work led by the amazing Michele Bevilacqua!

So excited about this! Tune in this Friday at 3 PM CEST...

We reframe #MT evaluation as sequence tagging in our #WMT2022 paper: check MaTESe, an interpretable metric that tags error spans in translations! 📜 statmt.org/wmt22/pdf/2022… By Stefano perrella Lorenzo Proietti Alessandro Scirè Niccolò Campolungo Roberto Navigli Babelscape European Research Council (ERC) #EMNLP2022

Wow! Today it has >4k monthly downloads on Hugging Face 🤗

Exciting strides in text summarization with LLMs 🚀but verifying their factual accuracy is still an open challenge 🤔 We introduce FENICE, a factuality-oriented metric for summarization with a strong focus on interpretability🔍arxiv.org/abs/2403.02270 #NLProc #LLMs #Factuality

By leveraging NLI-based alignments, FENICE matches claims extracted from the summary to passages in the source document. This allows identifying the specific sections that support or contradict each claim, providing deeper insights into the factual accuracy of a generated summary

FENICE achieves state of the art performance on AGGREFACT, the de-facto benchmark for this task, and excels in our set of human annotations for long-form summarization evaluation. arxiv.org/abs/2403.02270 Joint work between SapienzaNLP and Babelscape Karim Kamal Roberto Navigli

FENICE: Factuality Evaluation of summarization based on Natural language Inference and Claim Extraction - Findings Poster Session 4 📅14th August @ 12:15 🗣️ Alessandro Scirè Karim Kamal