- Original Pdf: pdf
- Abstract: Pretrained language models attract lots of attentions, and they take advantage of the two-stages training process: pretraining on huge corpus and finetuning on specific tasks. Thereinto, BERT (Devlin et al., 2019) is a Transformer (Vaswani et al., 2017) based model and has been the state-of-the-art for many kinds of Nature Language Processing (NLP) tasks. However, BERT cannot take text longer than the maximum length as input since the maximum length is predefined during pretraining. When we apply BERT to long text tasks, e.g., document-level text summarization: 1) Truncating inputs by the maximum sequence length will decrease performance, since the model cannot capture long dependency and global information ranging the whole document. 2) Extending the maximum length requires re-pretraining which will cost a mass of time and computing resources. What's even worse is that the computational complexity will increase quadratically with the length, which will result in an unacceptable training time. To resolve these problems, we propose to apply Transformer to only model local dependency and recurrently capture long dependency by inserting multi-channel LSTM into each layer of BERT. The proposed model is named as BERT-AL (BERT for Arbitrarily Long Document Understanding) and it can accept arbitrarily long input without re-pretraining from scratch. We demonstrate BERT-AL's effectiveness on text summarization by conducting experiments on the CNN/Daily Mail dataset. Furthermore, our method can be adapted to other Transformer based models, e.g., XLNet (Yang et al., 2019) and RoBERTa (Liu et al., 2019), for various NLP tasks with long text.