Massimo Nicosia

📄 Our new EMNLP paper is on arXiv! 📄 1⃣ Train an mT5 filler model to reconstruct full parses from English utterances + parse signatures 2⃣ Run it on translations and parse signatures to obtain high quality i18n synthetic data! More here:👉 arxiv.org/abs/2109.04319 👈 @Google

Great to be in #singapore for #emnlp2023 with two contributions: check out the XTREME-UP and mmT5 papers! XTREME-UP: arxiv.org/abs/2305.11938 mmT5: arxiv.org/abs/2305.14224 #EMNLP2023 #NLProc #google @GoogleDeepMind @GoogleAI @Google

Attending @Google @GoogleDeepMind google network and social event was great too as I had the opportunity to network with various AI practitioners.

#Indaba2023 is a wrap! Six days of packed workshops, presentations, and panels. With an attendance of over 800 AI practitioners 32 African countries. But the main highlight was the people and hearing about their journeys. Thanks to the organizing committee for a great experience

Thanks for organizing the sessions @2ayasalama ! It was great meeting you and chatting with all the participants 😃 #Indaba2023 #DLI_23 #Yɛbetumi

I had a great time meeting folks from all over Africa (plus several Google DeepMind colleagues from UK/US...) at the #Indaba2023

At Deep Learning Indaba! See you at the Google booths or at Tuesday's Breakfast Mentorship Session. Thursday aft I'll present a "Request For Plot", a short project that students can complete to start a collab with Google researchers. LMK if you are here and would like to chat!

We are excited to release QAmeleon dataset, a multilingual QA dataset with 47,000+ LLM-generated QA pairs in 8 languages! github.com/google-researc… Training multilingual QA models with this bridges ~60% of the gap between an En-only baseline and a fully-supervised upper bound.

🚨🚨🚨 Check out this exciting work led by team mate @PfeiffJo, a new modular model that is great at cross-lingual transfer! Happy to have contributed to this project! 😊

XTREME-UP: A User-Centric Scarce-Data Benchmark for Under-Represented Languages propose XTREME-UP, a benchmark defined by: its focus on the scarce-data scenario rather than zero-shot; its focus on user-centric tasks -- tasks with broad adoption by speakers of high-resource languages; and its focus on under-represented languages where this scarce-data scenario tends to be most realistic. XTREME-UP evaluates the capabilities of language models across 88 under-represented languages over 9 key user-centric technologies including ASR, OCR, MT, and information access tasks that are of general utility. We create new datasets for OCR, autocomplete, semantic parsing, and transliteration, and build on and refine existing datasets for other tasks. XTREME-UP provides methodology for evaluating many modeling scenarios including text-only, multi-modal (vision, audio, and text),supervised parameter tuning, and in-context learning paper page: huggingface.co/papers/2305.11…

I'm very happy that MasakhaNER github.com/masakhane-io/m… by @MasakhaneNLP is part of XTREME-UP, you can now evaluate on several low-resource languages in a more realistic setting. Thank you @GoogleAI team for this new benchmark.

🚨 I'm really excited to release XTREME-UP, a new benchmark focusing on under-represented languages. We emphasize a realistic evaluation setting including new and existing user-centric tasks and realistic data sizes beyond the few-shot setting.

Our new work is finally out! 🥳 XTREME-UP, A User-Centric Scarce-Data Benchmark for Under-Represented Languages. Proud of scaling NLP to more under-represented languages while collaborating with so many talented folks! Find more info here: ▶️ github.com/google-researc…

Thanks AK for sharing! TL; DR: We propose MatCha🍵, visual language pretraining by (1) mapping charts to their underlying data tables and rendering code (2) decoding answers of math questions rendered as images.

📍🧵🚨 QA on plots & charts is a complex task requiring sophisticated reasoning - our visual language models struggle with this. LLMs are super strong reasoners - but they only work for text. What do we do? We translate plots & charts to text so LLM can understand!

In the Massively Multilingual NLU 2022 workshop (mmnlu-22.github.io), @maxnicosia showcases the Translate and Fill approach (aclanthology.org/2021.findings-…). For multilingual text generation, they get better results with ByT5 models over mT5 for smaller models. #EMNLP2022

@PfeiffJo @Google Welcome to the team Jonas! 😉🥳

I’m happy to announce that I have started at @Google Research in Zürich as a Research Scientist 😍🥳

@AmazonScience @nopper @emnlpmeeting Thank you!

@maxnicosia @nopper @emnlpmeeting Congrats both!

🥇 Happy to announce that @nopper and I won the Zero-Shot task of the MMNLU-22 Challenge organized by Amazon! Our recipe? Translate-and-Fill synthetic data and ByT5. 📄TaF paper: aclanthology.org/2021.findings-… 🏆 MMNLU-22: mmnlu-22.github.io @emnlpmeeting @GoogleAI @AmazonScience

@OG_rohan @PMinervini @nopper @emnlpmeeting @GoogleAI @AmazonScience @heymaxichrome Thank you Rohan! We'll look into that. Please note that our system was trained without gold data from non-en languages. If your goal is to serve the most accurate parser, you may be better off training ByT5 huggingface.co/google/byt5-ba… on all the available gold data from Massive.

@maxnicosia @PMinervini @nopper @emnlpmeeting @GoogleAI @AmazonScience @heymaxichrome great achievement 🔥. Do you and @nopper plan on releasing the model publicly anytime soon? Would love to host it on the ML API Marketplace that I am building.

@Kuzeko @nopper @emnlpmeeting @GoogleAI @AmazonScience Thank you Matteo!

@maxnicosia @nopper @emnlpmeeting @GoogleAI @AmazonScience Congratulations!!