English | [简体中文](./README_zh.md) ## _ERNIE-Doc_: A Retrospective Long-Document Modeling Transformer - [Framework](#framework) - [Pre-trained Models](#Pre-trained-Models) - [Fine-tuning Tasks](#Fine-tuning-Tasks) * [Language Modeling](#Language-Modeling) * [Long-Text Classification](#Long-Text-Classification) * [Question Answering](#Question-Answering) * [Information Extraction](#Information-Extraction) * [Semantic Matching](#Semantic-Matching) - [Usage](#Usage) * [Install Paddle](#Install-PaddlePaddle) * [Fine-tuning](#Fine-tuning) - [Citation](#Citation) For technical description of the algorithm, please see our paper: >[_**ERNIE-Doc: A Retrospective Long-Document Modeling Transformer**_](https://arxiv.org/abs/2012.15688) > >Siyu Ding\*, Junyuan Shang\*, Shuohuan Wang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang (\* : equal contribution) > >Preprint December 2020 > >Accepted by **ACL-2021** ![ERNIE-Doc](https://img.shields.io/badge/Pretraining-Long%20Document%20Modeling-green) ![paper](https://img.shields.io/badge/Paper-ACL2021-yellow) --- **ERNIE-Doc is a document-level language pretraining model**. Two well-designed techniques, namely the **retrospective feed mechanism** and the **enhanced recurrence mechanism**, enable ERNIE-Doc, which has a much longer effective context length, to capture the contextual information of a complete document. ERNIE-Doc improved the state-of-the-art language modeling result of perplexity to 16.8 on WikiText-103. Moreover, it outperformed competitive pretraining models by a large margin on most language understanding tasks, such as text classification, question answering, information extraction and semantic matching. ## Framework We proposed three novel methods to enhance the long document modeling ability of Transformers: - **Retrospective Feed Mechanism**: Inspired by the human reading behavior of skimming a document first and then looking back upon it attentively, we design a retrospective feed mechanism in which segments from a document are fed twice as input. As a result, each segment in the retrospective phase could explicitly fuse the semantic information of the entire document learned in the skimming phase, which prevents context fragmentation. - **Enhanced Recurrence Mechansim**, a drop-in replacement for a Recurrence Transformer (like Transformer-XL), by changing the shifting-one-layer-downwards recurrence to the same-layer recurrence. In this manner, the maximum effective context length can be expanded, and past higher-level representations can be exploited to enrich future lower-level representations. - **Segment-reordering Objective**, a document-aware task of predicting the correct order of the permuted set of segments of a document, to model the relationship among segments directly. This allows ERNIE-Doc to build full document representations for prediction. ![framework](.meta/framework.png) Illustrations of ERNIE-Doc and Recurrence Transformers, where models with three layers take as input a long document which is sliced into four segments. ## Pre-trained Models We release the checkpoints for **ERNIE-Doc _base_en/zh_** and **ERNIE-Doc _large_en_** model。 - [**ERNIE-Doc _base_en_**](https://ernie-github.cdn.bcebos.com/model-ernie-doc-base-en.tar.gz) (_12-layer, 768-hidden, 12-heads_) - [**ERNIE-Doc _base_zh_**](https://ernie-github.cdn.bcebos.com/model-ernie-doc-base-zh.tar.gz) (_12-layer, 768-hidden, 12-heads_) - [**ERNIE-Doc _large_en_**](https://ernie-github.cdn.bcebos.com/model-ernie-doc-large-en.tar.gz) (_24-layer, 1024-hidden, 16-heads_) ## Fine-tuning Tasks We compare the performance of [ERNIE-Doc](https://arxiv.org/abs/2012.15688) with the existing SOTA pre-training models (such as [Longformer](https://arxiv.org/abs/2004.05150), [BigBird](https://arxiv.org/abs/2007.14062), [ETC](https://arxiv.org/abs/2004.08483) and [ERNIE2.0](https://arxiv.org/abs/1907.12412)) for language modeling (**_WikiText-103_**) and document-level natural language understanding tasks, including long-text classification (**_IMDB_**, **_HYP_**, **_THUCNews_**, **_IFLYTEK_**), question answering (**_TriviaQA_**, **_HotpotQA_**, **_DRCD_**, **_CMRC2018_**, **_DuReader_**, **_C3_**), information extraction (**_OpenKPE_**) and semantic matching (**_CAIL2019-SCM_**). ### Language Modeling - [WikiText-103](https://arxiv.org/abs/1609.07843) | Model | Param. | PPL | |--------------------------|:--------:|:------:| | _Results of base models_ | | | | LSTM | - | 48.7 | | LSTM+Neural cache | - | 40.8 | | GCNN-14 | - | 37.2 | | QRNN | 151M | 33.0 | | Transformer-XL Base | 151M | 24.0 | | SegaTransformer-XL Base | 151M | 22.5 | | **ERNIE-Doc** Base | 151M | **21.0** | | _Results of large models_ | | | | Adaptive Input | 247M | 18.7 | | Transformer-XL Large | 247M | 18.3 | | Compressive Transformer | 247M | 17.1 | | SegaTransformer-XL Large | 247M | 17.1 | | **ERNIE-Doc** Large | 247M | **16.8** | ### Long-Text Classification - [IMDB reviews](http://ai.stanford.edu/~amaas/data/sentiment/index.html) | Models | Acc. | F1 | |-----------------|:----:|:----:| | RoBERTa | 95.3 | 95.0 | | Longformer | 95.7 | - | | BigBird | - | 95.2 | | **ERNIE-Doc** Base | **96.1** | **96.1** | | XLNet-Large | 96.8 | - | - | | **ERNIE-Doc** Large | **97.1** | **97.1** | - [Hyperpartisan News Dection](https://pan.webis.de/semeval19/semeval19-web/) | Models | F1 | |-----------------|:----:| | RoBERTa | 87.8 | | Longformer | 94.8 | | BigBird | 92.2 | | **ERNIE-Doc** Base | **96.3** | | **ERNIE-Doc** Large | **96.6** | - [THUCNews(THU)](http://thuctc.thunlp.org/)、[IFLYTEK(IFK)](https://arxiv.org/abs/2004.05986) | Models | THU | THU | IFK | |-----------------|:--------:|:--------:|:--------:| | | Acc. | Acc. | Acc. | | | Dev | Test | Dev | | BERT | 97.7 | 97.3 | 60.3 | | BERT-wwm-ext | 97.6 | 97.6 | 59.4 | | RoBERTa-wwm-ext | - | - | 60.3 | | ERNIE 1.0 | 97.7 | 97.3 | 59.0 | | ERNIE 2.0 | 98.0 | 97.5 | 61.7 | | **ERNIE-Doc** | **98.3** | **97.7** | **62.4** | ### Question Answering - [TriviaQA](http://nlp.cs.washington.edu/triviaqa/) on dev-set | Models | F1 | |-----------------|:----:| | RoBERTa | 74.3 | | Longformer | 75.2 | | BigBird | 79.5 | | **ERNIE-Doc** Base | **80.1** | | Longformer Large | 77.8 | | BigBird Large | - | | **ERNIE-Doc** Large | **82.5** | - [HotpotQA](https://hotpotqa.github.io/) on dev-set | Models | Span-F1 | Supp.-F1 | Joint-F1 | |-----------------|:----:|:----:|:----:| | RoBERTa | 73.5 | 83.4 | 63.5 | | Longformer | 74.3 | 84.4 | 64.4 | | BigBird | 75.5 | **87.1** | 67.8 | | **ERNIE-Doc** Base | **79.4** | 86.3 | **70.5** | | Longformer Large | 81.0 | 85.8 | 71.4 | | BigBird Large | 81.3 | **89.4** | - | | **ERNIE-Doc** Large | **82.2** | 87.6 | **73.7** | - [DRCD](https://arxiv.org/abs/1806.00920), [CMRC2018](https://arxiv.org/abs/1810.07366), [DuReader](https://arxiv.org/abs/1711.05073), [C3](https://arxiv.org/abs/1904.09679) | Models | DRCD | DRCD | CMRC2018 | DuReader | C3 | C3 | |-----------------|---------------|---------------|---------------|---------------|----------|----------| | | dev | test | dev | dev | dev | test | | | EM/F1 | EM/F1 | EM/F1 | EM/F1 | Acc. | Acc. | | BERT | 85.7/91.6 | 84.9/90.9 | 66.3/85.9 | 59.5/73.1 | 65.7 | 64.5 | | BERT-wwm-ext | 85.0/91.2 | 83.6/90.4 | 67.1/85.7 | -/- | 67.8 | 68.5 | | RoBERTa-wwm-ext | 86.6/92.5 | 85.2/92.0 | 67.4/87.2 | -/- | 67.1 | 66.5 | | MacBERT | 88.3/93.5 | 87.9/93.2 | 69.5/87.7 | -/- | - | - | | XLNet-zh | 83.2/92.0 | 82.8/91.8 | 63.0/85.9 | -/- | - | - | | ERNIE 1.0 | 84.6/90.9 | 84.0/90.5 | 65.1/85.1 | 57.9/72/1 | 65.5 | 64.1 | | ERNIE 2.0 | 88.5/93.8 | 88.0/93.4 | 69.1/88.6 | 61.3/74.9 | 72.3 | 73.2 | | **ERNIE-Doc** | **90.5/95.2** | **90.5/95.1** | **76.1/91.6** | **65.8/77.9** | **76.5** | **76.5** | ### Information Extraction - [Open Domain Web Keyphrase Extraction](https://www.aclweb.org/anthology/D19-1521/) | Models | F1@1 | F1@3 | F1@5 | |-----------|:----:|:----:|:----:| | BLING-KPE | 26.7 | 29.2 | 20.9 | | JointKPE | 39.1 | 39.8 | 33.8 | | ETC | - | 40.2 | - | | ERNIE-Doc | **40.2** | **40.5** | **34.4** | ### Semantic Matching - [CAIL2019-SCM](https://arxiv.org/abs/1911.08962) | Models | Dev (Acc.) | Test (Acc.) | |-----------|:-------------:|:-------------:| | BERT | 61.9 | 67.3 | | ERNIE 2.0 | 64.9 | 67.9 | | ERNIE-Doc | **65.6** | **68.8** | ## Usage ### Install PaddlePaddle This code base has been tested with Paddle (version>=1.8) with Python3. Other dependency of ERNIE-GEN is listed in `requirements.txt`, you can install it by ```script pip install -r requirements.txt ``` ### Fine-tuning We release the finetuning code for English and Chinese classification tasks and Chinese Question Answers Tasks. For example, you can finetune **ERNIE-Doc** base model on IMDB and IFLYTEK dataset by ```shell sh script/run_imdb.sh sh script/run_iflytek.sh sh script/run_dureader.sh ``` [Preprocessing code for IMDB dataset](./ernie_doc/data/imdb/README.md) The log of training and the evaluation results are in `log/job.log.0`. **Notice**: The actual total batch size is equal to `configured batch size * number of used gpus`. ## Citation You can cite the paper as below: ``` @article{ding2020ernie, title={ERNIE-DOC: The Retrospective Long-Document Modeling Transformer}, author={Ding, Siyu and Shang, Junyuan and Wang, Shuohuan and Sun, Yu and Tian, Hao and Wu, Hua and Wang, Haifeng}, journal={arXiv preprint arXiv:2012.15688}, year={2020} } ```