metadata

title: Mirror
emoji: 🪞
colorFrom: blue
colorTo: yellow
sdk: docker
pinned: true
license: apache-2.0

🪞 Mirror: A Universal Framework for Various Information Extraction Tasks

Image generated by DALLE 3
[Paper] | [Demo]
📃 Our paper has been accepted to EMNLP23 main conference, check it out!

😎: This is the official implementation of 🪞Mirror which supports almost all the Information Extraction tasks.

The name, Mirror, comes from the classical story Snow White and the Seven Dwarfs, where a magic mirror knows everything in the world. We aim to build such a powerful tool for the IE community.

🔥 Supported Tasks

Named Entity Recognition
Entity Relationship Extraction (Triplet Extraction)
Event Extraction
Aspect-based Sentiment Analysis
Multi-span Extraction (e.g. Discontinuous NER)
N-ary Extraction (e.g. Hyper Relation Extraction)
Extractive Machine Reading Comprehension (MRC) and Question Answering
Classification & Multi-choice MRC

🌴 Dependencies

Python>=3.10

pip install -r requirements.txt

🚀 QuickStart

Pretrained Model Weights & Datasets

Download the pretrained model weights & datasets from [OSF] .

No worries, it's an anonymous link just for double blind peer reviewing.

Pretraining

Download and unzip the pretraining corpus into resources/Mirror/v1.4_sampled_v3/merged/all_excluded
Start to run

CUDA_VISIBLE_DEVICES=0 rex train -m src.task -dc conf/Pretrain_excluded.yaml

Fine-tuning

⚠️ Due to data license constraints, some datasets are unavailable to provide directly (e.g. ACE04, ACE05).

Download and unzip the pretraining corpus into resources/Mirror/v1.4_sampled_v3/merged/all_excluded
Download and unzip the fine-tuning datasets into resources/Mirror/uie/
Start to fine-tuning

# UIE tasks
CUDA_VISIBLE_DEVICES=0 bash scripts/single_task_wPTAllExcluded_wInstruction/run1.sh
CUDA_VISIBLE_DEVICES=1 bash scripts/single_task_wPTAllExcluded_wInstruction/run2.sh
CUDA_VISIBLE_DEVICES=2 bash scripts/single_task_wPTAllExcluded_wInstruction/run3.sh
CUDA_VISIBLE_DEVICES=3 bash scripts/single_task_wPTAllExcluded_wInstruction/run4.sh
# Multi-span and N-ary extraction
CUDA_VISIBLE_DEVICES=4 bash scripts/single_task_wPTAllExcluded_wInstruction/run_new_tasks.sh
# GLUE datasets
CUDA_VISIBLE_DEVICES=5 bash scripts/single_task_wPTAllExcluded_wInstruction/glue.sh

Analysis Experiments

Few-shot experiments : scripts/run_fewshot.sh. Collecting results: python mirror_fewshot_outputs/get_avg_results.py
Mirror w/ PT w/o Inst. : scripts/single_task_wPTAllExcluded_woInstruction
Mirror w/o PT w/ Inst. : scripts/single_task_wo_pretrain
Mirror w/o PT w/o Inst. : scripts/single_task_wo_pretrain_wo_instruction

Evaluation

Change task_dir and data_pairs you want to evaluate. The default setting is to get results of Mirror_direct on all downstream tasks.
CUDA_VISIBLE_DEVICES=0 python -m src.eval

Demo

Download and unzip the pretrained task dump into mirror_outputs/Mirror_Pretrain_AllExcluded_2
Try our demo:

CUDA_VISIBLE_DEVICES=0 python -m src.app.api_backend

📋 Citation

@misc{zhu_mirror_2023,
  shorttitle = {Mirror},
  title = {Mirror: A Universal Framework for Various Information Extraction Tasks},
  author = {Zhu, Tong and Ren, Junfei and Yu, Zijian and Wu, Mengsong and Zhang, Guoliang and Qu, Xiaoye and Chen, Wenliang and Wang, Zhefeng and Huai, Baoxing and Zhang, Min},
  url = {http://arxiv.org/abs/2311.05419},
  doi = {10.48550/arXiv.2311.05419},
  urldate = {2023-11-10},
  publisher = {arXiv},
  month = nov,
  year = {2023},
  note = {arXiv:2311.05419 [cs]},
  keywords = {Computer Science - Artificial Intelligence, Computer Science - Computation and Language},
}

🛣️ Roadmap

Convert current model into Huggingface version, supporting loading from transformers like other newly released LLMs.
Remove Background area, merge TL, TP into a single T token
Add more task data: keyword extraction, coreference resolution, FrameNet, WikiNER, T-Rex relation extraction dataset, etc.
Pre-train on all the data (including benchmarks) to build a nice out-of-the-box toolkit for universal IE.

💌 Yours sincerely

This project is licensed under Apache-2.0. We hope you enjoy it ~

Mirror Team w/ 💖