Source code and dataset for ACL 2019 paper "ERNIE: Enhanced Language Representation with Informative Entities"

Related tags

Text Data & NLPERNIE
Overview

ERNIE

Source code and dataset for "ERNIE: Enhanced Language Representation with Informative Entities"

Reqirements:

  • Pytorch>=0.4.1
  • Python3
  • tqdm
  • boto3
  • requests
  • apex (If you want to use fp16, you should make sure the commit is 79ad5a88e91434312b43b4a89d66226be5f2cc98.)

Prepare Pre-train Data

Run the following command to create training instances.

  # Download Wikidump
  wget https://dumps.wikimedia.org/enwiki/latest/enwiki-latest-pages-articles.xml.bz2
  # Download anchor2id
  wget -c https://cloud.tsinghua.edu.cn/f/1c956ed796cb4d788646/?dl=1 -O anchor2id.txt
  # WikiExtractor
  python3 pretrain_data/WikiExtractor.py enwiki-latest-pages-articles.xml.bz2 -o pretrain_data/output -l --min_text_length 100 --filter_disambig_pages -it abbr,b,big --processes 4
  # Modify anchors with 4 processes
  python3 pretrain_data/extract.py 4
  # Preprocess with 4 processes
  python3 pretrain_data/create_ids.py 4
  # create instances
  python3 pretrain_data/create_insts.py 4
  # merge
  python3 code/merge.py

If you want to get anchor2id by yourself, run the following code(this will take about half a day) after python3 pretrain_data/extract.py 4

  # extract anchors
  python3 pretrain_data/utils.py get_anchors
  # query Mediawiki api using anchor link to get wikibase item id. For more details, see https://en.wikipedia.org/w/api.php?action=help.
  python3 pretrain_data/create_anchors.py 256 
  # aggregate anchors 
  python3 pretrain_data/utils.py agg_anchors

Run the following command to pretrain:

  python3 code/run_pretrain.py --do_train --data_dir pretrain_data/merge --bert_model ernie_base --output_dir pretrain_out/ --task_name pretrain --fp16 --max_seq_length 256

We use 8 NVIDIA-2080Ti to pre-train our model and there are 32 instances in each GPU. It takes nearly one day to finish the training (1 epoch is enough).

Pre-trained Model

Download pre-trained knowledge embedding from Google Drive/Tsinghua Cloud and extract it.

tar -xvzf kg_embed.tar.gz

Download pre-trained ERNIE from Google Drive/Tsinghua Cloud and extract it.

tar -xvzf ernie_base.tar.gz

Note that the extraction may be not completed in Windows.

Fine-tune

As most datasets except FewRel don't have entity annotations, we use TAGME to extract the entity mentions in the sentences and link them to their corresponding entitoes in KGs. We provide the annotated datasets Google Drive/Tsinghua Cloud.

tar -xvzf data.tar.gz

In the root directory of the project, run the following codes to fine-tune ERNIE on different datasets.

FewRel:

python3 code/run_fewrel.py   --do_train   --do_lower_case   --data_dir data/fewrel/   --ernie_model ernie_base   --max_seq_length 256   --train_batch_size 32   --learning_rate 2e-5   --num_train_epochs 10   --output_dir output_fewrel   --fp16   --loss_scale 128
# evaluate
python3 code/eval_fewrel.py   --do_eval   --do_lower_case   --data_dir data/fewrel/   --ernie_model ernie_base   --max_seq_length 256   --train_batch_size 32   --learning_rate 2e-5   --num_train_epochs 10   --output_dir output_fewrel   --fp16   --loss_scale 128

TACRED:

python3 code/run_tacred.py   --do_train   --do_lower_case   --data_dir data/tacred   --ernie_model ernie_base   --max_seq_length 256   --train_batch_size 32   --learning_rate 2e-5   --num_train_epochs 4.0   --output_dir output_tacred   --fp16   --loss_scale 128 --threshold 0.4
# evaluate
python3 code/eval_tacred.py   --do_eval   --do_lower_case   --data_dir data/tacred   --ernie_model ernie_base   --max_seq_length 256   --train_batch_size 32   --learning_rate 2e-5   --num_train_epochs 4.0   --output_dir output_tacred   --fp16   --loss_scale 128 --threshold 0.4

FIGER:

python3 code/run_typing.py    --do_train   --do_lower_case   --data_dir data/FIGER   --ernie_model ernie_base   --max_seq_length 256   --train_batch_size 2048   --learning_rate 2e-5   --num_train_epochs 3.0   --output_dir output_figer  --gradient_accumulation_steps 32 --threshold 0.3 --fp16 --loss_scale 128 --warmup_proportion 0.2
# evaluate
python3 code/eval_figer.py    --do_eval   --do_lower_case   --data_dir data/FIGER   --ernie_model ernie_base   --max_seq_length 256   --train_batch_size 2048   --learning_rate 2e-5   --num_train_epochs 3.0   --output_dir output_figer  --gradient_accumulation_steps 32 --threshold 0.3 --fp16 --loss_scale 128 --warmup_proportion 0.2

OpenEntity:

python3 code/run_typing.py    --do_train   --do_lower_case   --data_dir data/OpenEntity   --ernie_model ernie_base   --max_seq_length 128   --train_batch_size 16   --learning_rate 2e-5   --num_train_epochs 10.0   --output_dir output_open --threshold 0.3 --fp16 --loss_scale 128
# evaluate
python3 code/eval_typing.py   --do_eval   --do_lower_case   --data_dir data/OpenEntity   --ernie_model ernie_base   --max_seq_length 128   --train_batch_size 16   --learning_rate 2e-5   --num_train_epochs 10.0   --output_dir output_open --threshold 0.3 --fp16 --loss_scale 128

Some code is modified from the pytorch-pretrained-BERT. You can find the explanation of most parameters in pytorch-pretrained-BERT.

As the annotations given by TAGME have confidence score, we use --threshlod to set the lowest confidence score and choose the annotations whose scores are higher than --threshold. In this experiment, the value is usually 0.3 or 0.4.

The script for the evaluation of relation classification just gives the accuracy score. For the macro/micro metrics, you should use code/score.py which is from tacred repo.

python3 code/score.py gold_file pred_file

You can find gold_file and pred_file on each checkpoint in the output folder (--output_dir).

New Tasks:

If you want to use ERNIE in new tasks, you should follow these steps:

  • Use an entity-linking tool like TAGME to extract the entities in the text
  • Look for the Wikidata ID of the extracted entities
  • Take the text and entities sequence as input data

Here is a quick-start example (code/example.py) using ERNIE for Masked Language Model. We show how to annotate the given sentence with TAGME and build the input data for ERNIE. Note that it will take some time (around 5 mins) to load the model.

# If you haven't installed tagme
pip install tagme
# Run example
python3 code/example.py

Cite

If you use the code, please cite this paper:

@inproceedings{zhang2019ernie,
  title={{ERNIE}: Enhanced Language Representation with Informative Entities},
  author={Zhang, Zhengyan and Han, Xu and Liu, Zhiyuan and Jiang, Xin and Sun, Maosong and Liu, Qun},
  booktitle={Proceedings of ACL 2019},
  year={2019}
}
Owner
THUNLP
Natural Language Processing Lab at Tsinghua University
THUNLP
Python library to make development of portfolio analysis faster and easier

Trafalgar Python library to make development of portfolio analysis faster and easier Installation 🔥 For the moment, Trafalgar is still in beta develo

Santosh Passoubady 641 Jan 01, 2023
Longformer: The Long-Document Transformer

Longformer Longformer and LongformerEncoderDecoder (LED) are pretrained transformer models for long documents. ***** New December 1st, 2020: Longforme

AI2 1.6k Dec 29, 2022
中文医疗信息处理基准CBLUE: A Chinese Biomedical LanguageUnderstanding Evaluation Benchmark

English | 中文说明 CBLUE AI (Artificial Intelligence) is playing an indispensabe role in the biomedical field, helping improve medical technology. For fur

452 Dec 30, 2022
Universal Adversarial Triggers for Attacking and Analyzing NLP (EMNLP 2019)

Universal Adversarial Triggers for Attacking and Analyzing NLP This is the official code for the EMNLP 2019 paper, Universal Adversarial Triggers for

Eric Wallace 248 Dec 17, 2022
Minimal GUI for accessing the Watson Text to Speech service.

Description Minimal graphical application for accessing the Watson Text to Speech service. Requirements Python 3 plus all dependencies listed in requi

Moritz Maxeiner 1 Oct 22, 2021
Text Classification in Turkish Texts with Bert

You can watch the details of the project on my youtube channel Project Interface Project Second Interface Goal= Correctly guessing the classification

42 Dec 31, 2022
PyTorch Language Model for 1-Billion Word (LM1B / GBW) Dataset

PyTorch Large-Scale Language Model A Large-Scale PyTorch Language Model trained on the 1-Billion Word (LM1B) / (GBW) dataset Latest Results 39.98 Perp

Ryan Spring 114 Nov 04, 2022
Reproduction process of BERT on SST2 dataset

BERT-SST2-Prod Reproduction process of BERT on SST2 dataset 安装说明 下载代码库 git clone https://github.com/JunnYu/BERT-SST2-Prod 进入文件夹,安装requirements pip ins

yujun 1 Nov 18, 2021
The ability of computer software to identify words and phrases in spoken language and convert them to human-readable text

speech-recognition-py Speech recognition is the ability of computer software to identify words and phrases in spoken language and convert them to huma

Deepangshi 1 Apr 03, 2022
[Preprint] Escaping the Big Data Paradigm with Compact Transformers, 2021

Compact Transformers Preprint Link: Escaping the Big Data Paradigm with Compact Transformers By Ali Hassani[1]*, Steven Walton[1]*, Nikhil Shah[1], Ab

SHI Lab 367 Dec 31, 2022
Lightweight utility tools for the detection of multiple spellings, meanings, and language-specific terminology in British and American English

Breame ( British English and American English) Breame is a lightweight Python package with a number of utility tools to aid in the detection of words

Charles 8 Oct 10, 2022
Yomichad - a Japanese pop-up dictionary that can display readings and English definitions of Japanese words

Yomichad is a Japanese pop-up dictionary that can display readings and English definitions of Japanese words, kanji, and optionally named entities. It is similar to yomichan, 10ten, and rikaikun in s

Jonas Belouadi 7 Nov 07, 2022
基于Transformer的单模型、多尺度的VAE模型

UniVAE 基于Transformer的单模型、多尺度的VAE模型 介绍 https://kexue.fm/archives/8475 依赖 需要大于0.10.6版本的bert4keras(当前还没有推到pypi上,可以直接从GitHub上clone最新版)。 引用 @misc{univae,

苏剑林(Jianlin Su) 49 Aug 24, 2022
A2T: Towards Improving Adversarial Training of NLP Models (EMNLP 2021 Findings)

A2T: Towards Improving Adversarial Training of NLP Models This is the source code for the EMNLP 2021 (Findings) paper "Towards Improving Adversarial T

QData 17 Oct 15, 2022
Code for text augmentation method leveraging large-scale language models

HyperMix Code for our paper GPT3Mix and conducting classification experiments using GPT-3 prompt-based data augmentation. Getting Started Installing P

NAVER AI 47 Dec 20, 2022
Source code of the "Graph-Bert: Only Attention is Needed for Learning Graph Representations" paper

Graph-Bert Source code of "Graph-Bert: Only Attention is Needed for Learning Graph Representations". Please check the script.py as the entry point. We

14 Mar 25, 2022
LUKE -- Language Understanding with Knowledge-based Embeddings

LUKE (Language Understanding with Knowledge-based Embeddings) is a new pre-trained contextualized representation of words and entities based on transf

Studio Ousia 587 Dec 30, 2022
Creating an LSTM model to generate music

Music-Generation Creating an LSTM model to generate music music-generator Used to create basic sin wave sounds music-ai Contains the functions to conv

Jerin Joseph 2 Dec 02, 2021
Code for the Findings of NAACL 2022(Long Paper): AdapterBias: Parameter-efficient Token-dependent Representation Shift for Adapters in NLP Tasks

AdapterBias: Parameter-efficient Token-dependent Representation Shift for Adapters in NLP Tasks arXiv link: upcoming To be published in Findings of NA

Allen 16 Nov 12, 2022
BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents

BROS (BERT Relying On Spatiality) is a pre-trained language model focusing on text and layout for better key information extraction from documents. Given the OCR results of the document image, which

Clova AI Research 94 Dec 30, 2022