Named Entity Recognition with Tensorflow

This repo implements a NER model using Tensorflow (LSTM + CRF + chars embeddings). It includes modifications by CEA LIST for its use inside LIMA.

Versions : Tensoflow 1.3 and Python 3.5

Check the blog post

Task

Given a sentence, give a tag to each word. A classical application is Named Entity Recognition (NER). Here is an example

John   lives in New   York
B-PER  O     O  B-LOC I-LOC

Model

Similar to Lample et al. and Ma and Hovy.

• concatenate final states of a bi-lstm on character embeddings to get a character-based representation of each word
• concatenate this representation to a standard word vector representation (GloVe here)
• run a bi-lstm on each sentence to extract contextual representation of each word
• decode with a linear chain CRF

Getting started

We need to download a dataset in any language you want and pre-trained embeddings (not necessarily).

For English, GloVe embeddings are the best.

To download the GloVe vectors :

make glove

Alternatively, you can download them manually here and update the glove_filename entry in config.py. You can also choose not to load pretrained word vectors by changing the entry use_pretrained to False in model/config.py.

To build the training data, train and evaluate the model with

make run

Details

Here is the breakdown of the commands executed in make run:

The language used by default is English.

1. [DO NOT MISS THIS STEP] Build vocab from the data and extract trimmed glove vectors according to the config in model/config.py.

python build_data.py --lang eng/fr

2. Train the model with

python train.py --lang eng/fr

3. Evaluate and interact with the model with

python evaluate.py --lang eng/fr

4. Export the model with

python freezeGraph.py --lang eng/fr

5. Evaluate with C++ API

python testAPIC++.py --lang eng/fr

Data iterators and utils are in model/data_utils.py and the model with training/test procedures is in model/ner_model.py

Training Data

The training data must be in the following format named IOB (Inside-Output-Beginning), identical to the CoNLL'03 dataset [1].

It is recommended to use IOBES annotations, performances are better with this format.

State-of-the-art performance (F1 score between 90 and 91) have been reached training on the English corpora CoNLL'03.

Results with French and English are avalaible in results folder.

I use English CoNLL'03 dataset + GloVe embeddings (d-300) and WikiNER [2] dataset (aij-wikiner-fr-wp3) + fastText embeddings (d-300) [3] for French.

In config.py, all are specified.

Once you have produced your data files, change the parameters in config.py like

# dataset
dev_filename = "data/coNLL/eng/eng.testa.iob"
test_filename = "data/coNLL/eng/eng.testb.iob"
train_filename = "data/coNLL/eng/eng.train.iob"

If you want to use a new language, you have to precise it in config.py like

if(self.language=='yourlanguage'):
    # outputs
    self.dir_output = '...'

    # embeddings
    self.dim_word = '...'
    self.dim_char = '...'

    # embeddings files
    self.filename_glove = "...".format(self.dim_word)

    # trimmed embeddings (created from glove_filename with build_data.py)
    self.filename_trimmed = "...".format(self.dim_word)
    self.use_pretrained = True/False

    self.dir_resources="data/format_used/yourlanguage/" format_used={IOB1,IOB2,IOBES,...}

    # dataset
    self.filename_dev = "data/format_used/yourlanguage/yourlanguage.testa"
    self.filename_test = "data/format_used/yourlanguage/yourlanguage.testb"
    self.filename_train = "data/format_used/yourlanguage/yourlanguage.train"

    # vocab (created from dataset with build_data.py)
    self.filename_words = "data/format_used/yourlanguage/words.txt"
    self.filename_tags = "data/format_used/yourlanguage/tags.txt"
    self.filename_chars = "data/format_used/yourlanguage/chars.txt"

To create French corpora based on WikiNER, you need to follow these instructions :

data/system2conll.pl aijwikinerenwp3.bz2
python parse_fr_data.py

To create English corpora, you need to follow the README provided by coNLL folder.

To compare the old process unit from the new one, you can use the script Python compare.py

To install files in LIMA you have to precise the paths in apic++/moduleNER/CMakeLists.txt

License

This project is licensed under the terms of the apache 2.0 license (as Tensorflow and derivatives). If used for research, citation would be appreciated.

Citations

[1] Sang, Erik F. Tjong Kim, et Fien De Meulder. « Introduction to the CoNLL-2003 Shared Task: Language-Independent Named Entity Recognition ». arXiv:cs/0306050, 12 juin 2003. http://arxiv.org/abs/cs/0306050.

[2] Nothman, Joe, Nicky Ringland, et Will Radford. « Learning multilingual named entity recognition from Wikipedia », 2012. https://ac.els-cdn.com/S0004370212000276/1-s2.0-S0004370212000276-main.pdf?_tid=d2d8c47c-e003-11e7-a9d8-00000aab0f6b&acdnat=1513169448_9f6447041adb80408317a8d2019ed899.

[3] https://github.com/facebookresearch/fastText/blob/master/pretrained-vectors.md