Neural ParsCit v1.0.7 (Training Fix)

Release Notes

1. Fixed issue when training with train, dev and test that might have non-complete labels.
2. Replace hard-casting to float32 in certain part of the code to use theano.config.floatX.

Neural ParsCit v1.0.6 (Bug Fix)

Due to the restoration of training pipeline, the tag_to_id argument is mandatory in prepare_dataset.

A small hack is introduced to handle the non-requirement of tag_to_id during inference in run.py

Neural ParsCit v1.0.5 (Training Fix)

Release Notes

• Restore the commented out code during training and fixed missing tag in in data in training.
• Use sklearn for model evaluation (model performance, confusion matrix)
• Python 3 support for training code. Thanks @nsorros.

Notes

• Python 3 support is fairly weak; the webserver will not run in Python 3 due to the use of unicode instead of str. The next major release in PyTorch will be only Python 3.6+ compatible.

Neural ParsCit v1.0.4 (REST API)

Release Notes

• freq and vectors.kv are deprecated for support in this release
• REST API implementation (/parscit/parse and /parscit/parse/batch). Please refer to the /docs for the API documentation.

Notes

• When using the Docker image for REST API, you will need to pip install -r requirements/prod.txt

Neural ParsCit v1.0.3 (Optimisation)

Release Notes

• freq is not longer required as the entire word vectors provided by WING can be loaded with lower peak memory usage than the v1.0.2
• Added environment variable ENVIRONMENT to the Docker container. You may want to rebuild your image.
• If the <UNK> is not provided in the word embeddings, the embedding layer will issue a warning about higher memory consumption as it will add it into the word embeddings in memory (< 7.5 GB)
• Otherwise, if <UNK> exists in the word embeddings, it will be more memory efficient (< 4.5 GB). You will need to get the word embeddings with <UNK> from README
• Introduced model evaluation (skipped in CI due to insufficient memory)
  • v1.0.2 (macro f1: 0.984859387481312, micro f1: 0.9935205183585313)
  • v1.0.3 (macro f1: 0.9843960859239398, micro f1: 0.9933405327573794)

Important Notes

• This version is the last supported version for Theano-based Neural ParsCit unless major issues are discovered
• The model will be ported to PyTorch in v1.1 and will support Python 3.5+ only
• The word embeddings without <UNK> and freq files will not be downloadable in v1.0.4
• The existing files: word vectors (vectors.tar.gz and freq) will be removed from WING website in v1.0.4

Neural ParsCit v1.0.2 (Optimisation)

Release Notes

• The memory footprint has improved further. A Docker container with 8 GB should be able to build the model
• gensim, Theano and numpy have been upgraded to the latest version (you will need to rebuild your Docker image)
• Word vectors are loaded with KeyedVectors because of API deprecation. (Please read the important notes below and README with regards to this as there is breaking change with regards to how to build the model)

Important Notes

• As the word vectors are loaded lazily to reduce memory footprint, you will need to download the new word vectors from WING website and follow the instruction on README

Neural ParsCit v1.0.1 (Pristine)

Release Notes

This release improves the documentation and usability of the model with the original code.

• Include the Docker image definition
• Configure Theano in the container to use 32-bit FP and OpenBLAS
• Provide requirements.txt to freeze dependencies

In order to run in the container, a minimum of 12 GB RAM is required.