❤️
A Toolkit for Document-level Event Extraction with & without Triggers
Hi, there
This project aims at building a universal toolkit for extracting events automatically from documents
Currently, this repo contains PTPCG, Doc2EDAG and GIT models, and these models are all designed for document-level event extraction without triggers. Here are some basic descriptions to help you understand the characteristics of each model:
- PTPCG is a fast and lightweight model which takes only 3.6% GPU training resources than GIT, and it is 8.5x faster when inference. Besides, PTPCG is better than GIT and Doc2EDAG on o2o (one instance per doc) docs. Overall metrics scores are higher than Doc2EDAG and competitive to GIT. We tested this model on the LIC'21 information extraction shared task and won top5 prize
🏆 (team: 广告位招租). Availability are confirmed. Supplements are included here (including detailed examples, the BK algorithm, hyperparameters and additional experiment results).
- GIT is the SOTA model (Doc2EDAG modification), which shows the great power on o2m (multi instances with the same event type per doc) and m2m (multi instances with multi event types per doc) docs. GIT is slower than Doc2EDAG and need more resources to train and inference.
- Doc2EDAG is an auto-regressive model with good compatibilities on DocEE without triggers and is a widely used baseline for DocEE. This repo is developed based on Doc2EDAG.
⚙️
Installation
Make sure you have the following dependencies installed.
- Python 3.7.7
- torch==1.5.1 # should be OK with higher torch version
- pytorch-mcrf==0.0.3 # for MaskedCRF
- gpu-watchmen==0.3.8 # if you wanna wait for a vacant GPU via
gpu-watchmen - loguru==0.5.3
- matplotlib==3.3.0
- numpy==1.19.0
- transformers==4.9.1
- dgl-cu92==0.6.1 # find a version that is compatable with your CUDA version
- tqdm==4.53.0
- networkx==2.4
- tensorboard==2.4.1
# don't forget to install the dee package
$ git clone https://github.com/Spico197/DocEE.git
$ pip install -e .
# or install directly from git
$ pip install git+https://github.com/Spico197/DocEE.git
🚀
Quick Start
💾
Data Preprocessing
# ChFinAnn
## You can download Data.zip from the original repo: https://github.com/dolphin-zs/Doc2EDAG
$ unzip Data.zip
$ cd Data
# generate data with doc type (o2o, o2m, m2m) for better evaluation
$ python stat.py
# DuEE-fin
## If you want to win the test, you should check the codes and make further modifications,
## since each role may refer to multiple entities in DuEE-fin.
## Our PTPCG can help with this situation, all you need is to check the data preprocessing
## and check `predict_span_role()` method in `event_table.py`.
## We **do not** perform such magic tricks in the paper to make fair comparisons with Doc2EDAG and GIT.
$ # downloading datasets from https://aistudio.baidu.com/aistudio/competition/detail/65
$ cd Data/DuEEData # paste train.json and dev.json into Data/DuEEData folder and run:
$ python build_data.py
📋
To Reproduce Results in Paper
Doc2EDAG and GIT are already integrated in this repo, and more models are planned to be added.
If you want to reproduce the PTPCG results, or run other trials, please follow the instructions below.
Before running any bash script, please ensure bert_model has been correctly set.
Tip: At least 4 * NVIDIA V100 GPU (at least 16GB) cards are required to run Doc2EDAG models.
# run on ChFinAnn dataset
$ nohup bash scripts/run_doc2edag.sh 1>Logs/Doc2EDAG_reproduction.log 2>&1 &
$ tail -f Logs/Doc2EDAG_reproduction.log
# run on DuEE-fin dataset without trigger
$ nohup bash scripts/run_doc2edag_dueefin.sh.sh 1>Logs/Doc2EDAG_DuEE_fin.log 2>&1 &
$ tail -f Logs/Doc2EDAG_DuEE_fin.log
# run on DuEE-fin dataset with trigger
$ nohup bash scripts/run_doc2edag_dueefin_withtgg.sh 1>Logs/Doc2EDAG_DuEE_fin_with_trigger.log 2>&1 &
$ tail -f Logs/Doc2EDAG_DuEE_fin_with_trigger.log
Tip: At least 4 * NVIDIA V100 GPU (32GB) cards are required to run GIT models.
# run on ChFinAnn dataset
$ nohup bash scripts/run_git.sh 1>Logs/GIT_reproduction.log 2>&1 &
$ tail -f Logs/GIT_reproduction.log
# run on DuEE-fin dataset without trigger
$ nohup bash scripts/run_git_dueefin.sh 1>Logs/GIT_DuEE_fin.log 2>&1 &
$ tail -f Logs/GIT_DuEE_fin.log
# run on DuEE-fin dataset with trigger
$ nohup bash scripts/run_git_dueefin_withtgg.sh 1>Logs/GIT_DuEE_fin_with_trigger.log 2>&1 &
$ tail -f Logs/GIT_DuEE_fin_with_trigger.log
Tip: At least 1 * 1080Ti (at least 9GB) card is required to run PTPCG.
Default: |R| = 1, which means only the first (pseudo) trigger is selected.
# run on ChFinAnn dataset (to reproduce |R|=1 results in Table 1 of the PTPCG paper)
$ nohup bash scripts/run_ptpcg.sh 1>Logs/PTPCG_R1_reproduction.log 2>&1 &
$ tail -f Logs/PTPCG_R1_reproduction.log
# run on DuEE-fin dataset without annotated trigger (to reproduce |R|=1, Tgg=× results in Table 3 of the PTPCG paper)
$ nohup bash scripts/run_ptpcg_dueefin.sh 1>Logs/PTPCG_P1-DuEE_fin.log 2>&1 &
$ tail -f Logs/PTPCG_P1-DuEE_fin.log
# run on DuEE-fin dataset with annotated trigger and without pseudo trigger (to reproduce |R|=0, Tgg=√ results in Table 3 of the PTPCG paper)
$ nohup bash scripts/run_ptpcg_dueefin_withtgg.sh 1>Logs/PTPCG_T1-DuEE_fin.log 2>&1 &
$ tail -f Logs/PTPCG_T1-DuEE_fin.log
# run on DuEE-fin dataset with annotated trigger and one pseudo trigger (to reproduce |R|=1, Tgg=√ results in Table 3 of the PTPCG paper)
$ nohup bash scripts/run_ptpcg_dueefin_withtgg_withptgg.sh 1>Logs/PTPCG_P1T1-DuEE_fin.log 2>&1 &
$ tail -f Logs/PTPCG_P1T1-DuEE_fin.log
| #PseudoTgg | Setting | Log | Task Dump |
|---|---|---|---|
| 1 | 189Cloud | 189Cloud | 189Cloud |
Explainations on PTPCG hyperparameters in the executable script:
# whether to use max clique decoding strategy, brute-force if set to False
max_clique_decode = True
# number of triggers when training, to make all arguments as pseudo triggers, set to higher numbers like `10`
num_triggers = 1
# number of triggers when evaluating, set to `-1` to make all arguments as pseudo triggers
eval_num_triggers = 1
# put additional pseudo triggers into the graph, make full use of the pseudo triggers
with_left_trigger = True
# make the trigger graph to be directed
directed_trigger_graph = True
# run mode is used in `dee/tasks/dee_task.py/DEETaskSetting`
run_mode = 'full'
# at least one combination (see paper for more information)
at_least_one_comb = True
# whether to include regex matched entities
include_complementary_ents = True
# event schemas, check `dee/event_types` for all support schemas
event_type_template = 'zheng2019_trigger_graph'
⚽
Find Pseudo Triggers
Please check Data/trigger.py for more details. In general, you should first convert your data into acceptable format (like typed_train.json after building ChFinAnn).
Then, you can run the command below to generate event schemas with pseudo triggers and importance scores:
$ cd Data
$ python trigger.py <max number of pseudo triggers>
📚
Instructions
deehas evoluted to a toolkit package, make sure to install the package first:pip install -e .- Please change the path to BERT to load the tokenizer.
- To run on ChFinAnn dataset, you should generate
typed_(train|dev|test).jsonfiles first viacd Data && python stat.pyafterData.zipfile unzipped into theDatafolder. - It's not DDP model by default. If you want to train across different devices, add a
--parallel_decorateflag afterpython run_dee_task.py. - Comments starting with
tzhuare added by Tong Zhu to help understanding the codes, not written in the original Doc2EDAG repo. - For trials on DuEE-fin dataset, if you want to submit generated files to online platform, check the
dueefin_post_process.pyto make further post process to meet the format requirments. - I had tried lots of wasted models, so there were redundancies. For better understanding the codes and get rid of any potential distractions, I delete them from this repo. There may be some other redundancies and you may find there are unused methods or models, feel free to touch me and make the repo cleaner and nicer together~ Btw, there may be some issues if some files are removed directly. Feel free to reach me by openning an issue or email. I check the GitHub site messages everyday regularly and emails are received instantly during weekdays.
🙋
FAQ
- Q: What's the evluation strategy to calculate the final micro-F1 scores?
- A: Micro-F1 scores are calculated by counting the final number of event role predictions' TP , FP and FNs
- Q: What is
teacher_probdoing ?- A: It's used in the scheduled sampling strategy, indicating the probability to use the
gold_span. Ifteacher_prob == 0.7, then there is 70% probability to usegold_spanduring training.teacher_probwill decrease during training.
- A: It's used in the scheduled sampling strategy, indicating the probability to use the
- What's GreedyDec?
- A: Greedy decoding is a prediction generation strategy. We can fill in the event table by finding the first corresponding entity for each field (argument role). That's why it's called a
Greedymethod.
- A: Greedy decoding is a prediction generation strategy. We can fill in the event table by finding the first corresponding entity for each field (argument role). That's why it's called a
- Q: How to make predictions and get readable results with a trained model?
- A: Such inference interface is provided in
dee/tasks/dee_task.py/DEETask.predict_one()(Convenient online serving interface).
- A: Such inference interface is provided in
- Q: What is
o2o,o2mandm2m?- A: They are abbreviations for
one-type one-instance per doc,one-type with multiple instances per docandmultiple types per doc.
- A: They are abbreviations for
📜
Citation
This work has not been published yet, please cite the arXiv preview version first
@misc{zhu-et-al-2021-ptpcg,
title={Efficient Document-level Event Extraction via Pseudo-Trigger-aware Pruned Complete Graph},
author={Tong Zhu and Xiaoye Qu and Wenliang Chen and Zhefeng Wang and Baoxing Huai and Nicholas Jing Yuan and Min Zhang},
year={2021},
eprint={2112.06013},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
🔑
Licence
MIT Licence
🤘
Furthermore
This repo is still under development. If you find any bugs, don't hesitate to drop us an issue.
Thanks~
图1
146 Nov 29, 2022
69 Nov 15, 2022
109 Dec 28, 2022
32 Dec 15, 2022
47 Nov 22, 2022
5 Oct 30, 2022
3 Mar 9, 2022
49 Nov 21, 2022
153 Jan 7, 2023
58 Dec 21, 2022
39 Dec 16, 2022
146 Nov 29, 2022
18 Dec 11, 2022
4 Dec 31, 2022
270 Dec 31, 2022
1.2k Jan 4, 2023
13 Jan 6, 2023