Deep Sequence Classification

Generic library for training models for deep neural networks for text sequence classification tasks.

Usage:

• Create a json file config.json (default name) using the template in config.json.sample and specify the parameters for your training.

• Best strategy is to save the training and test files in vector format in advance and then give their paths in data_vectors parameter in the file.

• Train a model:

python model.py --config config_multitask.json --verbose 1

• Resume training from saved weights:

python model.py --config config_multitask.json --verbose 1 --weights output/models/model_multi_brnn_multitask_h2-45.h5 --base_epochs 45

Preprocessing:

• Currently, we support the preprocessing for the following file formats:

<ROOT><DOC>
<DOCNO> DOCUMENT 1 </DOCNO>
     For <TIME TYPE="DATE:DATE">six years</TIME> , <ENTITY TYPE="PERSON">Shubhanshu A. B. Mishra</ENTITY> has made several programming projects after being inspired by <ENTITY TYPE="PERSON">Linus Torvalds</ENTITY>, a very renowned programmer.
</DOC></ROOT>

• Each file can contain multiple DOCNO.
• The dir structure consists of many folders of data split for cross validation. It is as follows:

data/
data/CV_files
data/CV_files/1/file1.xml
data/CV_files/1/file2.xml
data/CV_files/1/file3.xml
...

data/CV_files/5/file1.xml
data/CV_files/5/file2.xml

Supports:

• Boundary and Category Detection
• Simple RNN and Bidirectional RNN
• Multi task sequence learning (Boundary + Category trained using same model)
• CNN + BRNN

Coming Up:

Use Cases:

• Named Entity Recognition
• POS Tagging
• Dependency Parsing

Author:

• Shubhanshu Mishra

Dependencies:

• Theano
• Keras
• BeautifulSoup (with lxml)
• numpy
• lxml (requires libxml2, libxslt and libxml2-dev)

Install theano and keras using the following commands:

pip install --user --upgrade --no-deps git+git://github.com/Theano/Theano.git
pip install --user --upgrade --no-deps git+git://github.com/fchollet/keras.git