Longformer: The Long-Document Transformer

10 Apr 2020  ·  Iz Beltagy, Matthew E. Peters, Arman Cohan ·

Transformer-based models are unable to process long sequences due to their self-attention operation, which scales quadratically with the sequence length. To address this limitation, we introduce the Longformer with an attention mechanism that scales linearly with sequence length, making it easy to process documents of thousands of tokens or longer. Longformer's attention mechanism is a drop-in replacement for the standard self-attention and combines a local windowed attention with a task motivated global attention. Following prior work on long-sequence transformers, we evaluate Longformer on character-level language modeling and achieve state-of-the-art results on text8 and enwik8. In contrast to most prior work, we also pretrain Longformer and finetune it on a variety of downstream tasks. Our pretrained Longformer consistently outperforms RoBERTa on long document tasks and sets new state-of-the-art results on WikiHop and TriviaQA. We finally introduce the Longformer-Encoder-Decoder (LED), a Longformer variant for supporting long document generative sequence-to-sequence tasks, and demonstrate its effectiveness on the arXiv summarization dataset.

PDF Abstract

Results from the Paper

Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Language Modelling enwik8 Longformer (30 layers, h=512) Bit per Character (BPC) 0.99 # 11
Number of params 102M # 13
Language Modelling enwik8 Longformer (12 layers, h=512) Bit per Character (BPC) 1.00 # 17
Number of params 41M # 27
Language Modelling Hutter Prize Longformer Large Bit per Character (BPC) 0.99 # 4
Number of params 102M # 4
Language Modelling Hutter Prize Longformer Small Bit per Character (BPC) 1.00 # 6
Number of params 41M # 14
Question Answering WikiHop Longformer-large Test 81.9 # 2