Atlas: Few-shot Learning with Retrieval Augmented Language Models

Atlas: обучение в режиме few-shot с использованием языковых моделей, дополненных поиском
Patrick Lewis, Sebastian Riedel, Gautier Izacard, Édouard Grave, Armand Joulin, Timo Schick, María Lomelí, Fabio Petroni, Lucas Hosseini, Jane Dwivedi-Yu
2022-08-05

Natural Questionsdocument indexfew-shot learningknowledge-intensive tasksretrieval-augmented language models
Large language models have shown impressive few-shot results on a wide range of tasks. However, when knowledge is key for such results, as is the case for tasks such as question answering and fact checking, massive parameter counts to store knowledge seem to be needed. Retrieval augmented models are known to excel at knowledge intensive tasks without the need for as many parameters, but it is unclear whether they work in few-shot settings. In this work we present Atlas, a carefully designed and pre-trained retrieval augmented language model able to learn knowledge intensive tasks with very few training examples. We perform evaluations on a wide range of tasks, including MMLU, KILT and NaturalQuestions, and study the impact of the content of the document index, showing that it can easily be updated. Notably, Atlas reaches over 42% accuracy on Natural Questions using only 64 examples, outperforming a 540B parameters model by 3% despite having 50x fewer parameters.
1
Atlas is a carefully designed and pretrained retrieval-augmented language model that learns knowledge-intensive tasks from very few examples.
2
Atlas’s performance depends on document-index content, which can be easily updated to incorporate new or changed knowledge.
3
Evaluations cover diverse benchmarks, including MMLU, KILT, and NaturalQuestions, demonstrating Atlas’s broad few-shot capabilities.
4
On Natural Questions, Atlas exceeds 42% accuracy using only 64 examples and outperforms a 540B-parameter model by 3% with 50× fewer parameters.
5
The results show that retrieval augmentation can support strong few-shot knowledge-intensive learning without requiring massive parameter counts to store knowledge.

Atlas, a retrieval-augmented language model for few-shot learning on knowledge-intensive tasks

Few-shot learning of knowledge-intensive tasks and the effect of document-index content and updates

Publication Details
Publication Date
2022-08-05
Journal
Publisher
ISSN
Cited by
993
Access Type
Author Information
Authors
Patrick Lewis
Sebastian Riedel
Gautier Izacard
Édouard Grave
Armand Joulin
Timo Schick
María Lomelí
Fabio Petroni
Lucas Hosseini
Jane Dwivedi-Yu
Explore further
Open the scid.ai AI chat with a ready-made request: it will find papers on a similar topic and help build a literature review.
Find similar papers in the chat →
Make a presentation
100%