A CRF-BiLSTM model aims at quick and convenient benchmarking the performances of different word embeddings on your own corpus.

the objectives of this model are:

• Build a CRF-BiLSTM Network in Tensorflow with methods for easisly switching among different word embeddings (Word2vec, GloVe, Fasttext, ELMo, Flair and any combinations of them) while keep the same CRF-LSTM Network unchanged.

• Methods for easily gridsearch on the suitable parameters.

Python 3, TensorFlow 1.0+, Gensim, and Flair(optinal):

• if need to use "Contextual" embedding, Flair library(https://github.com/zalandoresearch/flair) is required.

1. Modify the directory to the Cropus and Configure the Hyper-parameter accordingly in config.py

    # embeddings_size
    dim_word = 300
    dim_char = 50
    #
    hidden_size_char = 64 # lstm on chars
    hidden_size_lstm = 128 # lstm on word embeddings

    # dataset
    path_data_root = 'data/CoNLL2003/'
    path_train = path_data_root +'eng.testa'
    path_eval = path_data_root +'eng.testa'
    path_test = path_data_root +'eng.testb'

2. Designate the embedding you want. Since different embeddings come with different file formats, this part maybe vary slightly accordding to the embedding you choose. there is a example for them in "How To Use.ipynb"

    # glove
	config = Config('glove')
	glove_file_path = 'data/glove/glove.6B.100d.txt'
	config.init_glove(glove_file_path)

    # fasttext
	config = Config('fasttext')
	command ='../fastText/fasttext'
	bin_file ='../fastText/data/cc.en.300.bin'
	config.init_fasttext(command, bin_file)

3. Parse the corpus and generate the "index" and "input". the following code will base on the vocabularies of embedding and corpus to generate the index for token/character/label and the mapping the each sentence into a sequence of index. this part also handle the sepcific configuration of model base on the corpus, like the number of kind of label, the number of unique character in corpus.

# parse the corpus and generate the input data
token2idx, char2idx, label2idx, lookup_table = get_idx(config)
train_x, train_y = get_inputs('train', token2idx, char2idx, label2idx, config)
eval_x, eval_y = get_inputs('eval', token2idx, char2idx, label2idx, config)
test_x, test_y = get_inputs('test', token2idx, char2idx, label2idx, config)

4. initial the model's graph and train/eval/test.

# initial the same NER model 
ner_model = Model(config)
ner_model.build_graph()
ner_model.initialize_session()

5. resutl: the F1 score based on the label will be print and detail of training processing can be find in "./output/log.log".

you could find more detail in "How To Use.ipynb"

This model is based on the following papers:

• Lample, Guillaume, et al. "Neural architectures for named entity recognition." arXiv preprint arXiv:1603.01360 (2016).
• Zhiheng Huang, et al. "Bidirectional LSTM-CRF Models for Sequence Tagging." arXiv preprint arXiv:1508.01991 (2015).