Bi-directional LSTM-CNNs-CRF for Italian Sequence Labeling

This is a sequence labeler that supports the following tasks:

• Named Enitity Recognition
• Parts of Speech Tagging
• SuperSense Tagging (use the Named Entity Recognition mode)

If you use this software please cite:

Pierpaolo Basile, Giovanni Semeraro, Pierluigi Cassotti. Bi-directional LSTM-CNNs-CRF for Italian Sequence Labeling, Fourth Italian Conference on Computational Linguistics (CLIC-it 2017), 2017.

Requirements

• Python 3.6
• Numpy
• Tensorflow https://www.tensorflow.org/install/
• Theano http://deeplearning.net/software/theano/install.html
• Keras https://github.com/phipleg/keras/tree/crf

How to use it

Run main.py
Main supports the following parameters:

• task Task type default='NER' choices=['NER','POS']
• tag_scheme Tag scheme default='IOBES' choices=['IOB2', 'IOBES']
• train_path Training file path
• dev_path Development file path
• test_path Test file path
• embed_char 1 for using character embeddings 0 otherwise default=1
• lowerword 1 for performing token lowercase default=1
• char_dim Character embeddings dimension default=30
• word_dim Word embeddings dimension default=300
• word2vecBINARY 1 if word2vec embeddings is binary 0 otherwise default=0
• label_column The column index of labels default=3
• word_column The column index of words default=0
• embedding Word embeddings tools default='word2vec' choices=['word2vec','glove','senna','random']
• dataset_encoding The dataset encoding format default='utf-8'
• embedding_encoding The embeddings encoding format default='utf-8'
• n_filters Filter number of convolutional networks default=30
• window_size Window size of convolutional networks default=3
• lstm_size The state size of Long Short Term Memory default=200
• epochs Number of epochs default=50
• batch_size Batch size default=10
• learning_alghoritm Optimization method choices=['Adadelta', 'Adagrad', 'SGD'] default='Adadelta'
• learning_rate The learning rate in case of SGD default=0.015
• decay The deacy rate in case of SGD default=0.05
• grad_clipping The clip value default=5.0
• feature_column The column index of the feature default=2
• features 1 for using features 0 otherwise default=0

main2.py Additional parameters:

• log Log file for PoS Tagging (Training set)
• log2 Log file for PoS Tagging (Development set)
• log3 Log file for PoS Tagging (Test set)
• log4 Log file for Sentiment (Training set)
• log5 Log file for Sentiment (Development set)
• log6 Log file for Sentiment (Test set)
• senti_train_path Training file path (sentiment)
• senti_test_path Test file path (sentiment)
• sentiment default=1 1 for performing sentiment analysis
• k_fold default=1 1 for performing KFOLD5
• k_n default=1 Specify the partion of training set to use in Development set