search

DFKI-NLP / sherlock

State-of-the-art Information Extraction
3 stars 1 forks source link

readme

🕵️ Sherlock

Code style: black

State-of-the-art Information Extraction

Tested with Python 3.8.

Download and install Sherlock:

git clone git@github.com:DFKI-NLP/sherlock.git
cd sherlock
pip install .

AllenNLP-based

The most straighforward approach to use and test a model is using allennlp:

To train, use the AllenNLP CLI. This requires you to setup a config file. This project includes two example configurations in the configs folder:

To train them you can train the models via:

# transformer
allennlp train configs/binary_rc/transformer.jsonnet -f -s <serialization dir>

# cnn
allennlp train configs/binary_rc/cnn.jsonnet -f -s <serialization dir>

To evaluate the model it is expected that you model.tar.gz file in the archive format from AllenNLP. Now you have two options:

# evaluation script
python ./scripts/eval_binary_relation_clf_allennlp.py \
  --eval_data_path <PATH TO EVAL DATA> \
  --test_data_path <PATH TO TEST DATA> \
  --do_eval \
  --do_predict \
  --eval_all_checkpoints \
  --per_gpu_batch_size 8 \
  --output_dir <SERIALIZATION DIR or PATH TO ARCHIVE> \
  --overwrite_results

# allennlp cli
allennlp evaluate <PATH TO ARCHIVE> <PATH TO EVAL DATA> \
  --cuda-device 0 \
  --batch-size 8 \

Configs

The crux of the configs lies in the dataset_reader and model section.

dataset_reader

The dataset_reader for AllenNLP is a patch-together of the dataset_reader from sherlock and the feature_converter from sherlock.

It inherits from allennlp.data.DatasetReader and its name ("type") is "sherlock". It accepts a dataset_reader_name, which must be a registered sherlock-dataset_reader and dataset_reader_kwargs to initialize the dataset_reader with correct arguments. The same happens for the feature_converter. Besides that, it takes the standart arguments that a AllenNLP-DatasetReader takes. For more details look into the documentation of the sherlock_dataset_reader.

model

The models directory contains the models which can be used as of now. Because of dependency-injection you can produce quite a lot with these models already: whereby the transformer model is limited to a certain type of (bert-like) transformers, the basic_relation_classifier can handle anything which fits into the schema of "embedder" -> "encoder" -> "classifier" (yes, theoretically transformer based models too).

For the transformers module it is important to give it the correct tokenizer keyword arguments, in this case additional_special_tokens, as it uses those to rescale its embedding dimension. There did not seem another generic and clean way to do this.

Huggingface-based

The original repo was written only with the transformers library support. Although it is possible to use transformers models via AllenNLP, Sherlock v2 still supports using the older codebase:

Named-entity recognition

For example, to train a NER model on the TACRED dataset:

./scripts/run_ner.py \
  --model_type bert \
  --model_name_or_path bert-base-uncased \
  --do_train \
  --do_eval \
  --do_predict \
  --evaluate_during_training \
  --eval_all_checkpoints \
  --do_lower_case \
  --data_dir <TACRED DIR> \
  --save_steps 8500 \
  --logging_steps 8500 \
  --max_seq_length 128 \
  --per_gpu_eval_batch_size=8 \
  --per_gpu_train_batch_size=8 \
  --learning_rate 2e-5 \
  --num_train_epochs 5.0 \
  --overwrite_cache \
  --overwrite_output_dir \
  --output_dir <OUTPUT DIR> \
  --cache_dir <CACHE DIR>

Relation classification

For example, to train a RC model on the TACRED dataset:

./scripts/run_binary_relation_clf.py \
  --model_type bert \
  --model_name_or_path bert-base-uncased \
  --do_train \
  --do_eval \
  --do_predict \
  --evaluate_during_training \
  --eval_all_checkpoints \
  --do_lower_case \
  --data_dir <TACRED DIR> \
  --save_steps 8500 \
  --logging_steps 8500 \
  --max_seq_length 128 \
  --per_gpu_eval_batch_size=8 \
  --per_gpu_train_batch_size=8 \
  --learning_rate 2e-5 \
  --num_train_epochs 5.0 \
  --overwrite_cache \
  --overwrite_output_dir \
  --entity_handling mark_entity_append_ner \
  --output_dir <OUTPUT DIR> \
  --cache_dir <CACHE DIR>

Tests

Tests are located in the directory tests. To run them, being in the root directory call:

py.test

or

pytest -sv

To call a specific test specify testfile and use -k flag:

pytest tests/feature_converters/token_classification_test.py -sv -k "truncate"

Installation issues