Longformer: The Long-Document Transformer
from KnowledgeDB.ai Hosted by KnowledgeDB
Ref: https://arxiv.org/abs/2004.05150 The paper introduces Longformer, a Transformer model designed to efficiently process long sequences. It addresses the quadratic complexity of standard self-attention by using a linear-scaling mechanism combining local windowed attention and task-motivated global attention. The authors demonstrate Longformer's effectiveness on character-level language modeling and various downstream tasks, achieving state-of-the-art results. Furthermore, they introduce Longformer-Encoder-Decoder (LED), a variant for sequence-to-sequence tasks, showcasing its success in long document summarization. The improved efficiency and performance are achieved through architectural modifications and strategic training procedures.
Embed this episode
No transcript yet.