Excited to share our #ecir2023 paper: A Unified Framework for Learned Sparse Retrieval!
Most LSRs differ in their 🔢 encoder, 📉 regularizer, & 👀 supervision. We test the effect of each in terms of effectiveness & efficiency.
💻https://t.co/S1O1VpU7h6
📃https://t.co/wmtsciXSwm
Context Embeddings for Efficient Answer Generation in RAG
Proposes an effective context compression method to reduce long context and speed up generation time in RAG systems.
The long contexts are compressed into to a small number of context embeddings which allow different compression rates that trade off decoding time for generation quality.
Quote: "We reduce inference time by up to 5.69 × and GFLOPs by up to 22 × while maintaining high performance."
RAG systems can take up a lot of computational effort that might lead to infeasible systems. So it's cool to see the use of context embeddings to mitigate this constraint. This is also promising to explore even larger amounts of contexts and large document collections.
1/4 I'm happy to announce that we at @irlab_amsterdam will organize the next edition of the European Summer School in Information Retrieval (ESSIR) @essir_eu on July 1-5 at the University of Amsterdam.
Early registration deadline: May 31.
More info: https://t.co/1tT25MZ4jp
Join us tomorrow at 5pm for Search Engines Amsterdam 🌊, with talks by @ArthurCamara and @arian_ask
about using LLMs for data augmentation in IR! 🎉
Register here: https://t.co/jOBKhjBmUx
Search Engines Amsterdam 🌊 is back! Next Friday, we host two great talks by @ArthurCamara and @arian_ask on using LLMs for data augmentation in IR.
Register now and join us next week: https://t.co/NA9SN4LiQz
📌 “Efficient Multi-Vector dense retrieval with Bit vectors” (EMVB), a novel framework for efficient query processing in multi-vector dense retrieval.
📌 First, EMVB employs a highly efficient pre-filtering step of passages using optimized bit vectors.
📌 Second, the computation of the centroid interaction happens column-wise, exploiting SIMD instructions, thus reducing its latency.
📌 Third, EMVB leverages Product Quantization (PQ) to reduce the memory footprint of storing vector representations while jointly allowing for fast late interaction
At 15:00 on 8th April, Andrew Yates @andrewyates from the University of Amsterdam @UvA_Amsterdam will give an #IRTalk talk entitled "An encoder-centric view of retrieval". Details at: https://t.co/V0encyoZOd
@GlasgowCS@ir_glasgow
From 2.5x (SPLADE) to 11.3x (ESPLADE) faster than the winning methods of the Big-ANN '23 Sparse Track. "Efficient Inverted Indexes for Approximate Retrieval over Learned Sparse Representations". LP @ #SIGIR2024. Joint work with @snbruch, @cosimorulli1, @rventurini_
📷 SEA February - Personalizing LLMs📷 Feb 23rd, 17:00 CET, in-person at Lab42 or online. Our next SEA meet-up will feature a talk by @HamedZamani (University of Massachusetts Amherst) on personalizing Large Language Models. https://t.co/SGgfYKuG3M
Happy to share that our paper "Translate-Distill: Learning Cross-Language Dense Retrieval by Translation and Distillation" is accepted at #ECIR24
It's our effort to bring Translate-Train and distillation together for training CLIR models!
https://t.co/lyUxlb3Mx8
I am looking for a PhD candidate to work on machine learning for search systems! 👩🎓👨🎓
We offer a fully funded position, freedom in your research direction, great working environment and very friendly colleagues.
You can apply here: https://t.co/ym7Wfjz6lL
RT is appreciated! 😄
Now, at EMNLP 2023, @maliannejadi is presenting our work titled 'Expand, Highlight, Generate: RL-driven Document Generation for Passage Reranking,' which has been accepted in the main track of EMNLP long papers and will be presented as a poster.
1/3
(1/2) 📢 SEA in September! Only 3 days left to this special edition of SEA with great line of speakers from industry and academia.
🖋️Sign up here:
https://t.co/U3tKCVxoJB