José G. C. de Souza (@accezz) header
José G. C. de Souza (@accezz) profile picture

PhD, Computer Science, Multilingual NLP and LLMs, Generative AI, Machine Translation, Machine Learning | Ex- @Unbabel, @eBay, PhD @UniTrento / @FBK_research

Lisbon, Portugal
Joined April 2008
Photo by José G. C. de Souza (@accezz)
José G. C. de Souza retweeted
🚀 Thrilled to present our QUEST poster at #NeurIPS2024 in Vancouver! 📅 Thursday, Dec 12th | 4:30-7:30 PM PST 📍 East Exhibit Hall A-C #3108 Looking forward to engaging discussions & connecting with fellow researchers! Stop by to learn more about our work. Joint work with @swetaagrawal20 @tozefarinhas @RicardoRei7 @accezz @andre_t_martins 🔗 Details: neurips.cc/virtual/2024/post… questdecoding.com
4
1
12
2,693
José G. C. de Souza retweeted
Today we release EuroLLM-9B: the best EU-made multilingual LLM of its size! Check the blog post for more info and results: huggingface.co/blog/eurollm-…. Stay tuned for the technical report and bigger and more powerful models!
4
24
9
85
14,011
José G. C. de Souza retweeted
Today we release the first EuroLLM paper and models: EuroLLM-1.7B and EuroLLM-1.7B-Instruct! The EuroLLM project will develop open-weight multilingual LLMs that understand and generate text in all official EU languages. Stay tuned for the bigger and stronger EuroLLMs (9B, 22B)!
3
18
7
76
13,438
José G. C. de Souza retweeted
I’m super proud to announce that our Tower 🗼 paper was accepted at COLM 24! To celebrate, we are releasing a new version of TowerInstruct based on Mistral 7B: it achieves similar performance to our 13B model, while being almost half the size! COLM Paper: openreview.net/pdf?id=EHPns3…
1
13
3
46
11,433
José G. C. de Souza retweeted
We offer this 1y postdoc to work with us on the @UTTERProject EU project on LLM-based agents ! Come work with us on 1 or several of these topics: i] managing uncertainty and ambiguity ii] improving the use of conversational context iii] ensuring the safety and alignment of LLMs.
9
13
1,573
José G. C. de Souza retweeted
The Chat Shared Task (WMT2024) is live! 💥💥 Happy to announce this year’s Chat Shared Task which aims to translate a corpus composed of genuine bilingual conversations from the customer support domain!
1
9
13
1,797
José G. C. de Souza retweeted
Steering Large Language Models for Machine Translation with Finetuning and In-Context (virtual poster) with @duartealves @nunonmg @JoãoAlves @josepombal @RicardoRei7 @accezz @andre_t_martins Do not hesitate to pass by the poster and say hi to Duarte !
1
5
24
1,800
José G. C. de Souza retweeted
Excited to share that we've won the prestigious #QualityEstimation shared task at #WMT23 for a 2nd year in a row!🎉 Our winning build was a LLM QE system with 11B parameters -- the largest #QE system ever built. 💪 Check out the win here >> hubs.li/Q021rMgZ0 @machtranslate
1
1
6
1,201
Old-ish news but this paper is *so good*. It does such a thorough and systematic comparison of, like, every conceivable setting you could use for n-best reranking or MBR. Extremely useful for figuring out what actually works.
Tired of beam search and all the heuristics needed to make it work well in MT? In our work accepted at #NAACL2022 (co-lead @tozefarinhas) we explore an alternative decoding method that leverages neural metrics to produce better translations! arxiv.org/abs/2205.00978 1/14
4
18
5,638
José G. C. de Souza retweeted
UTTER User Day Update! 🚀 Don't miss the panel on Multimodal LLMs and the Future of NLP Research, moderated by @andre_t_martins. The panelists include @alexandrabirch1, @carolscarton, @slippylolo, and @Turchi_Marco. Secure your spot here: lnkd.in/djBg7Ume #UTTERUserDay
10
1
9
2,960
José G. C. de Souza retweeted
Improving Machine Translation Evaluation with COMET v2.0 - Say Goodbye to Outdated Metrics! 🌟 With all the hype around GPT-4 is more important than ever to have reliable evaluation metrics and methodologies. That why we are excited to release COMET v2.0: resources.unbabel.com/r-d-bl…
4
15
1
39
3,034
José G. C. de Souza retweeted
Does your (EU) country attract or lose top researchers? AT, CH attract a lot. FR, DE, BE, ES are open places: as many leave as arrive. IE is a closed system. Italy, there is something very wrong: *very* few come, many go. Plot made with @ERC_Research data.
60
1,166
191
4,219
José G. C. de Souza retweeted
Para boa parte dos liberais brasileiros o ódio ao PT é maior do que o amor à liberdade. Minhas extensas críticas ao petismo são conhecidas. Só que Bolsonaro é o presidente mais antiliberal da Nova República — e seu chavismo verde oliva já está em curso. Textos sobre o tema no 🧵
139
643
69
5,528
José G. C. de Souza retweeted
Se, há alguns anos, Gregório Duvivier me dissesse que eu apoiaria o Lula no futuro, eu diria que ele enlouqueceu. Por outro lado, se eu dissesse pro Gregório que Lula seria novamente candidato, com o Alckmin de vice e apoiado pela Marina, ele diria que eu enlouqueci. (cont. 🧵)
12
7,114
1,665
68,483
José G. C. de Souza retweeted
I am hiring a PhD candidate to work on uncertainty-aware and explainable models of natural language generation. Fully funded for 4 years at University of Amsterdam. Deadline for applications: August 21. For details see probabll.github.io/jobs/
I am hiring a PhD candidate in ML for NLP to work on deep generative models of language. Fully funded 4-year position at the University of Amsterdam, Language Technology Lab. Apply by August 21. Details and link to call at vene.ro/jobs.html
3
117
5
226
José G. C. de Souza retweeted
Have some ideas on how to make MT model use contexts to improve translations? Test them at the WMT22 Chat Shared Task! wmt-chat-task.github.io We released baselines, and context already helps beating sentence-level models, but we are pretty sure you can do even better than us :)
The Chat Shared Task 2022 is live!💥💥 Happy to share that this year’s Chat Shared Task will use a new corpus composed by genuine bilingual conversations from the customer support domain! Detsets are out and baselines will soon be available as well! wmt-chat-task.github.io/ #wmt22
4
9
José G. C. de Souza retweeted
Our amazing research team won the Best Paper Award at @EAMT2022 for turning the high-quality #machinetranslation evaluation metric, COMET, into COMETinho. 🥰 Congrats @CatarinaFarinh7, @RicardoRei7, @accezz and everyone else involved in this project! 🥳
Replying to @EAMT_2024
The @Unbabel team won with the little metric that could COMETinho! #eamt2022 congrats!
4
20
José G. C. de Souza retweeted
Tired of beam search and all the heuristics needed to make it work well in MT? In our work accepted at #NAACL2022 (co-lead @tozefarinhas) we explore an alternative decoding method that leverages neural metrics to produce better translations! arxiv.org/abs/2205.00978 1/14
2
27
4
142