NLStruct

Natural language struturing library. Currently, it implements a nested NER model and a span classification model, but other algorithms might follow.

If you find this library useful in your research, please consider citing:

@phdthesis{wajsburt:tel-03624928,
  TITLE = {{Extraction and normalization of simple and structured entities in medical documents}},
  AUTHOR = {Wajsb{\"u}rt, Perceval},
  URL = {https://hal.archives-ouvertes.fr/tel-03624928},
  SCHOOL = {{Sorbonne Universit{\'e}}},
  YEAR = {2021},
  MONTH = Dec,
  KEYWORDS = {nlp ; structure ; extraction ; normalization ; clinical ; multilingual},
  TYPE = {Theses},
  PDF = {https://hal.archives-ouvertes.fr/tel-03624928/file/updated_phd_thesis_PW.pdf},
  HAL_ID = {tel-03624928},
  HAL_VERSION = {v1},
}

Features

• processes large documents seamlessly: it automatically handles tokenization and sentence splitting.
• do not train twice: an automatic caching mechanism detects when an experiment has already been run
• stop & resume with checkpoints
• easy import and export of data
• handles nested or overlapping entities
• multi-label classification of recognized entities
• strict or relaxed multi label end to end retrieval metrcis
• pretty logging with rich-logger
• heavily customizable, without config files (see train_ner.py)
• built on top of transformers and pytorch_lightning

Training models

How to train a NER model

from nlstruct.recipes import train_ner

model = train_ner(
    dataset={
        "train": "path to your train brat/standoff data",
        "val": 0.05,  # or path to your validation data
        # "test": # and optional path to your test data
    },
    finetune_bert=False,
    seed=42,
    bert_name="camembert/camembert-base",
    fasttext_file="",
    gpus=0,
    xp_name="my-xp",
    return_model=True,
)
model.save_pretrained("model.pt")

How to use it

from nlstruct import load_pretrained
from nlstruct.datasets import load_from_brat, export_to_brat

ner = load_pretrained("model.pt")
ner.eval()
ner.predict({"doc_id": "doc-0", "text": "Je lui prescris du lorazepam."})
# Out: 
# {'doc_id': 'doc-0',
#  'text': 'Je lui prescris du lorazepam.',
#  'entities': [{'entity_id': 0,
#    'label': ['substance'],
#    'attributes': [],
#    'fragments': [{'begin': 19,
#      'end': 28,
#      'label': 'substance',
#      'text': 'lorazepam'}],
#    'confidence': 0.9998705969553088}]}
export_to_brat(ner.predict(load_from_brat("path/to/brat/test")), filename_prefix="path/to/exported_brat")

How to train a NER model followed by a span classification model

from nlstruct.recipes import train_qualified_ner

model = train_qualified_ner(
    dataset={
        "train": "path to your train brat/standoff data",
        "val": 0.05,  # or path to your validation data
        # "test": # and optional path to your test data
    },
    finetune_bert=False,
    seed=42,
    bert_name="camembert/camembert-base",
    fasttext_file="",
    gpus=0,
    xp_name="my-xp",
    return_model=True,
)
model.save_pretrained("model.pt")

Ensembling

Easily ensemble multiple models (same architecture, different seeds):

model1 = load_pretrained("model-1.pt")
model2 = load_pretrained("model-2.pt")
model3 = load_pretrained("model-3.pt")
ensemble = model1.ensemble_with([model2, model3]).cuda()
export_to_brat(ensemble.predict(load_from_brat("path/to/brat/test")), filename_prefix="path/to/exported_brat")

Advanced use

Should you need to further configure the training of a model, please modify directly one of the recipes located in the recipes folder.

Install

This project is still under development and subject to changes.

pip install nlstruct==0.2.0