The MRQA 2019 workshop is approaching! Today we are pleased to release our findings from the shared task: arxiv.org/abs/1910.09753 Congrats to Baidu Research for the best system, D-Net! It achieved 72.5 F1 on the 12 held-out datasets, 10.7 higher than our BERT-large baseline .

Congrats to our #NLP and #PaddlePaddle team for ranking No.1 in the MRQA 2019 Shared Task with their D-NET framework, which improves the generalization of machine reading comprehension! View more at mrqa.github.io/assets/papers/… #emnlp2019

We are excited to co-host the Language and Intelligence Challenge again this year with a brand-new task setting and large-scale Chinese datasets! Get on board ASAP and win ￥350K cash. Registration closes on 5/12！ Learn more at lic2020.cipsc.org.cn #AI #ML #NLP

All researchers & developers are welcome to compete in the tasks of Machine Reading Comprehension, Conversational Recommendation, Semantic Parsing, Relation Extraction, and Event Extraction.

Introducing PLATO-2, Baidu's newest open-domain chatbot model that can talk about any topic in both Chinese & English. Our experiments show substantial improvements to PLATO-2 over other SOTA models in both Chinese & English evaluations. Read more at research.baidu.com/Blog/index-vie…

As Baidu #NLP turns 10 this year, we are thrilled to announce our new plan to build 100 Chinese NLP datasets in three years: Qian Yan, or "Thousands of Words" in Chinese. Read more👇 research.baidu.com/Blog/index-vie…

Our RocketQA introduced the techniques of cross-batch negative sampling, denoised hard negative sampling and data augmentation to improve dense passage retrieval for open-domain QA. Looking forward to seeing more revolution in this area. 🚀🚀🚀

a good read

Two "must read" papers in 2020 for those interested in Search. Embedding-based Retrieval in Facebook Search arxiv.org/abs/2006.11632 RocketQA: An Optimized Training Approach to Dense Passage Retrieval for Open-Domain Question Answering arxiv.org/abs/2010.08191

Reviewer #2: from joke to reality 😭

We often cite papers using arXiv info w/o noting that they are already PUBLISHED in confs like ACL 2025 and ICLR 2026. These incorrect bib entries are so annoying!😠 We introduce Rebiber, a tool to fix them using ACL Anthology and дблп - ☮ Stop the war!. 😆 Code: github.com/yuchenlin/rebi…

Registration for the 2021 Language & Intelligence Challenge is now open! This year's competition is based on LUGE, an open-source project of Chinese NLP benchmarks. Winners will share an RMB 300,000 prize pool, register by May 12th to enter! Learn more: lic2021.ccf.org.cn

#NAACL2021 We'll present our work on 🚀 RocketQA: An Optimized Training Approach to Dense Passage Retrieval for Open-Domain Question Answering at Session 16D (Wed, 7:40PM PDT). Come say hi & ask questions. The paper📄, code💾, model📂, and etc. are at: github.com/PaddlePaddle/R…

Our work latest work on dense retrieval - 🚀RocketQAv2 will be presented at #EMNLP2021! We propose a joint training approach for dense passage retrieval and passage re-ranking. Paper📄:arxiv.org/pdf/2110.07367… Code💾: github.com/paddlepaddle/R… Video🎥: underline.io/events/192/pos…

🚀 The #RocketQA pre-trained models can now be used directly from Jina Hub thanks to our partnership with RocketQA! Read our latest #blog to build a state-of-the-art QA application with Jina and RocketQA in just a few lines of code! 👉 bit.ly/3zl19bP

We provide an easy-to-use toolkit for running and fine-tuning dense retrievers, namely RocketQA. 🚀Pre-trained SOTA models 🚀1st open-source Chinese model 🚀Integration with Jina AI to help build an e2e question answering system GitHub: github.com/paddlepaddle/R…

Dense Text Retrieval based on Pretrained Language Models: A Survey Surveys recent advances in employing PLMs for dense text retrieval, including architecture design, training strategies, indexing, and pipeline optimization. 📝dl.acm.org/doi/10.1145/36… 👨🏽‍💻github.com/RUCAIBox/Dense…

IR needs new hills to climb. TREC yielded BM25 almost right away (year 3), MS MARCO yielded BERT cross-encoders, doc2query, ColBERT, DeepImpact/SPLADE, and MarginMSE/RocketQA almost immediately (years 1-2), and BEIR yielded a few such things in year 1. So, what's next?

We've just unveiled ERNIE 4.5 & X1! 🚀 As a deep-thinking reasoning model with multimodal capabilities, ERNIE X1 delivers performance on par with DeepSeek R1 at only half the price. Meanwhile, ERNIE 4.5 is our latest foundation model and new-generation native multimodal model.

Baidu is back!!! Two models and a new global web portal has been released today. Finally the phone verification is no longer needed!!! They do listen to the voice of developers. This model release includes both enhanced lightweighted thinking model ERNIE-4.5-21B-A3B-Thinking