REBEL: Relation Extraction By End-to-end Language generation

Last update: Jan 06, 2023

Related tags

Overview

REBEL: Relation Extraction By End-to-end Language generation

This is the repository for the Findings of EMNLP 2021 paper REBEL: Relation Extraction By End-to-end Language generation. We present a new linearization aproach and a reframing of Relation Extraction as a seq2seq task. The paper can be found here. If you use the code, please reference this work in your paper:

@inproceedings{huguet-cabot-navigli-2021-rebel,
title = "REBEL: Relation Extraction By End-to-end Language generation",
author = "Huguet Cabot, Pere-Llu{\'\i}s  and
  Navigli, Roberto",
booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2021",
month = nov,
year = "2021",
address = "Online and in the Barceló Bávaro Convention Centre, Punta Cana, Dominican Republic",
publisher = "Association for Computational Linguistics",
url = "https://github.com/Babelscape/rebel/blob/main/docs/EMNLP_2021_REBEL__Camera_Ready_.pdf",
}

Repo structure
| conf  # contains Hydra config files
  | data
  | model
  | train
  root.yaml  # hydra root config file
| data  # data
| datasets  # datasets scripts
| model # model files should be stored here
| src
  | pl_data_modules.py  # LightinigDataModule
  | pl_modules.py  # LightningModule
  | train.py  # main script for training the network
  | test.py  # main script for training the network
| README.md
| requirements.txt
| demo.py # Streamlit demo to try out the model
| setup.sh # environment setup script

Initialize environment

In order to set up the python interpreter we utilize conda , the script setup.sh creates a conda environment and install pytorch and the dependencies in "requirements.txt".

REBEL Model and Dataset

Model and Dataset files can be downloaded here:

https://osf.io/4x3r9/?view_only=87e7af84c0564bd1b3eadff23e4b7e54

Or you can directly use the model from Huggingface repo:

https://huggingface.co/Babelscape/rebel-large

", "").replace("
", "").replace("", "").split(): if token == "

": current = 't' if relation != '': triplets.append({'head': subject.strip(), 'type': relation.strip(),'tail': object_.strip()}) relation = '' subject = '' elif token == "

": current = 's' if relation != '': triplets.append({'head': subject.strip(), 'type': relation.strip(),'tail': object_.strip()}) object_ = '' elif token == "

": current = 'o' relation = '' else: if current == 't': subject += ' ' + token elif current == 's': object_ += ' ' + token elif current == 'o': relation += ' ' + token if subject != '' and relation != '' and object_ != '': triplets.append({'head': subject.strip(), 'type': relation.strip(),'tail': object_.strip()}) return triplets extracted_triplets = extract_triplets(extracted_text[0]) print(extracted_triplets) ">
from transformers import pipeline

triplet_extractor = pipeline('text2text-generation', model='Babelscape/rebel-large', tokenizer='Babelscape/rebel-large')

# We need to use the tokenizer manually since we need special tokens.
extracted_text = triplet_extractor.tokenizer.batch_decode(triplet_extractor("Punta Cana is a resort town in the municipality of Higuey, in La Altagracia Province, the eastern most province of the Dominican Republic", return_tensors=True, return_text=False)[0]["generated_token_ids"]["output_ids"])

print(extracted_text[0])

# Function to parse the generated text and extract the triplets
def extract_triplets(text):
    triplets = []
    relation, subject, relation, object_ = '', '', '', ''
    text = text.strip()
    current = 'x'
    for token in text.replace("", "").replace("
        
         "
        , "").replace("", "").split():
        if token == "
       
        "
       :
            current = 't'
            if relation != '':
                triplets.append({'head': subject.strip(), 'type': relation.strip(),'tail': object_.strip()})
                relation = ''
            subject = ''
        elif token == "
       
        "
       :
            current = 's'
            if relation != '':
                triplets.append({'head': subject.strip(), 'type': relation.strip(),'tail': object_.strip()})
            object_ = ''
        elif token == "
       
        "
       :
            current = 'o'
            relation = ''
        else:
            if current == 't':
                subject += ' ' + token
            elif current == 's':
                object_ += ' ' + token
            elif current == 'o':
                relation += ' ' + token
    if subject != '' and relation != '' and object_ != '':
        triplets.append({'head': subject.strip(), 'type': relation.strip(),'tail': object_.strip()})
    return triplets
extracted_triplets = extract_triplets(extracted_text[0])
print(extracted_triplets)

CROCODILE: automatiC RelatiOn extraCtiOn Dataset wIth nLi filtEring.

REBEL dataset can be recreated using our RE dataset creator CROCODILE

Training and testing

There are conf files to train and test each model. Within the src folder to train for CONLL04 for instance:

train.py model=rebel_model data=conll04_data train=conll04_train

Once the model is trained, the checkpoint can be evaluated by running:

test.py model=rebel_model data=conll04_data train=conll04_train do_predict=True checkpoint_path="path_to_checkpoint"

src/model_saving.py can be used to convert a pytorch lightning checkpoint into the hf transformers format for model and tokenizer.

DEMO

We suggest running the demo to test REBEL. Once the model files are unzipped in the model folder run:

streamlit run demo.py

And a demo will be available in the browser. It accepts free input as well as data from the sample file in data/rebel/

Datasets

TACRED is not freely avialable but instructions on how to create Re-TACRED from it can be found here.

For CONLL04 and ADE one can use the script from the SpERT github.

For NYT the dataset can be downloaded from Copy_RE github.

Finally the DocRED for RE can be downloaded at the JEREX github

REBEL: Relation Extraction By End-to-end Language generation

Related tags

Overview

REBEL: Relation Extraction By End-to-end Language generation

Initialize environment

REBEL Model and Dataset

CROCODILE: automatiC RelatiOn extraCtiOn Dataset wIth nLi filtEring.

Training and testing

DEMO

Datasets

Owner

Babelscape

Deep Learning for Natural Language Processing SS 2021 (TU Darmstadt)

Crawl & visualize ICLR papers and reviews

Real-time Object Detection for Streaming Perception, CVPR 2022

Tensorflow2 Keras-based Semantic Segmentation Models Implementation

AniGAN: Style-Guided Generative Adversarial Networks for Unsupervised Anime Face Generation

Python Wrapper for Embree

Evolutionary Scale Modeling (esm): Pretrained language models for proteins

Existing Literature about Machine Unlearning

JORLDY an open-source Reinforcement Learning (RL) framework provided by KakaoEnterprise

Human segmentation models, training/inference code, and trained weights, implemented in PyTorch

Official pytorch implementation of DeformSyncNet: Deformation Transfer via Synchronized Shape Deformation Spaces

Few-Shot Object Detection via Association and DIscrimination

Asterisk is a framework to generate high-quality training datasets at scale

The Codebase for Causal Distillation for Language Models.

Pytorch implementation of ICASSP 2022 paper Attention Probe: Vision Transformer Distillation in the Wild

This repo in the implementation of EMNLP'21 paper "SPARQLing Database Queries from Intermediate Question Decompositions" by Irina Saparina, Anton Osokin

Finetuner allows one to tune the weights of any deep neural network for better embeddings on search tasks

This repository contains the official implementation code of the paper Improving Multimodal Fusion with Hierarchical Mutual Information Maximization for Multimodal Sentiment Analysis, accepted at EMNLP 2021.

Finding Biological Plausibility for Adversarially Robust Features via Metameric Tasks

Reinfore learning tool box, contains trpo, a3c algorithm for continous action space