NERE

A Universal Toolkit for Neural Entity Relation Extraction.

Requirements

• python: 3.7.6
• pytorch: 1.8.1
• transformers: 4.2.2
• configargparse: 1.2.3
• bidict: 0.20.0
• fire: 0.3.1

Datasets

We provide scripts and instructions for processing two datasets (ACE2005,SciERC) and demo samples in the data/.

I/O

We decouple complex entity relation annotation into Field form. Each Field automatically constructs vocabulary and gets numerical input from the text.

Models

We jointly extract entities and relation by sharing parameters. The entity model is based on the sequence labeling framework, and the relation model is based on CNN network. This toolkit supports both the LSTM and pre-trained model as the sentence encoder. Besides, we implement two paradigm:

• Entity based: first detects entities then recognizes relations between entities.
• Span based: first detects entity spans then recognizes entities and relations.

Training

LSTM + Entity based

python entity_relation_extractor.py \
    --config_file config.yml \
    --save_dir ckpt/lstm_entity \
    --embedding_model word_char \
    --entity_model joint \
    --lstm_layers 1 \
    --lstm_hidden_unit_dims 256 \
    --epochs 300 \
    --device 0

LSTM + Span based

python entity_relation_extractor.py \
    --config_file config.yml \
    --save_dir ckpt/lstm_span \
    --embedding_model word_char \
    --entity_model pipeline \
    --lstm_layers 1 \
    --lstm_hidden_unit_dims 256 \
    --epochs 300 \
    --device 0

Pre-trained model + Entity based

python entity_relation_extractor.py \
    --config_file config.yml \
    --save_dir ckpt/ptm_entity \
    --embedding_model pretrained \
    --entity_model joint \
    --pretrained_model_name bert-base-uncased \
    --epochs 200 \
    --fine_tune \
    --device 0

Pre-trained model + Span based

python entity_relation_extractor.py \
    --config_file config.yml \
    --save_dir ckpt/ptm_span \
    --embedding_model pretrained \
    --entity_model pipeline \
    --pretrained_model_name bert-base-uncased \
    --epochs 200 \
    --fine_tune \
    --device 0

Inference

Note that save_dir should contain the trained best_model.

LSTM + Entity based

python entity_relation_extractor.py \
    --config_file config.yml \
    --save_dir ckpt/lstm_entity \
    --embedding_model word_char \
    --entity_model joint \
    --lstm_layers 1 \
    --lstm_hidden_unit_dims 256 \
    --device 0 \
    --log_file test.log \
    --test

LSTM + Span based

python entity_relation_extractor.py \
    --config_file config.yml \
    --save_dir ckpt/lstm_span \
    --embedding_model word_char \
    --entity_model pipeline \
    --lstm_layers 1 \
    --lstm_hidden_unit_dims 256 \
    --device 0 \
    --log_file test.log \
    --test

Pre-trained model + Entity based

python entity_relation_extractor.py \
    --config_file config.yml \
    --save_dir ckpt/ptm_entity \
    --embedding_model pretrained \
    --entity_model joint \
    --pretrained_model_name bert-base-uncased \
    --fine_tune \
    --device 0 \
    --log_file test.log \
    --test

Pre-trained model + Span based

python entity_relation_extractor.py \
    --config_file config.yml \
    --save_dir ckpt/ptm_span \
    --embedding_model pretrained \
    --entity_model pipeline \
    --pretrained_model_name bert-base-uncased \
    --fine_tune \
    --device 0 \
    --log_file test.log \
    --test

Evaluation

This toolkit supports different evaluation criteria:

• Entity: P/R/F1
• Span: P/R/F1
• Relation (relax): P/R/F1
• Relation (strict): P/R/F1

Notifications

• If OOM occurs, we suggest that reducing train_batch_size and increasing gradient_accumulation_steps (gradient_accumulation_steps is used to perform Gradient Accumulation).
• This toolkit currently does not support training on multi-GPUs.