Example code and data for paper Large-scale Exploration of Neural Relation Classification Architectures (EMNLP 2018).
MASS, version 1.0
, is a program that was developed with 3 main functions:
- Classify relation between entities pair in text documents
- Train new models with given corpora that follow the format
- Evaluate trained models with test dataset
We evaluated our model on 6 datasets, but in this example code, we included 2 typical datasets (one for generic data – semeval 2010, one for biomedical data – BC5).
This program was developed using Python version 3.5 and was tested on Linux and Windows system. We recommend using Anaconda 3 newest version for installing Python 3.5 as well as numpy, although you can install them by other means.
Other requirements:
- numpy
# Included in Anaconda package
- scipy
# Included in Anaconda package
- h5py
$ conda install -c anaconda h5py
- Tensorflow
$ pip install tensorflow # Python 3.n; CPU support
$ pip install tensorflow-gpu # Python 3.n; GPU support
If you are install tensorflow with GPU support, please follow the instructions on the official document to install other required libraries for your platform. Official document can be found at https://www.tensorflow.org/install/
- Keras
$ pip install keras
- sklearn
$ conda install -c anaconda scikit-learn
- NLTK
$ conda install -c anaconda nltk
You must download wordnet corpus using nltk.download()
- imbalanced-learn
$ conda install -c conda-forge imbalanced-learn
- fastText
$ git clone https://github.com/facebookresearch/fastText.git
$ cd fastText
$ pip install .
To download the pre-trained fastText model, go to the download page at: https://fasttext.cc/docs/en/pretrained-vectors.html. You should download English bin+text
zip file, extract the .zip file and then, put the wiki.en.bin
file at data/w2v_model
. Because of anonymous submission and review, we use generic fastText model for both BC5 and semeval 2010 datasets instead of using our trained fastText model on Biomedical text for BC5 dataset.
Before train and test the model, we must build some data:
- We pre-build trimmed embeddings before any run to boost up the speed of loading.
- We pre-build the dataset before feeding the data to the model. The datasets are then dumped to pickle file for later use.
To build dataset, run the following command:
$ python build_data.py
- The result contains: Two (2)
.npz
file, located atdata/w2v_model
- A pickle directory at
data/pickle
, containing four (4) pickle files.
Use one of two testing file (test_bc5.py
or test_semeval.py
) to train the model and evaluate on two benchmark datasets.
Commands:
$ python test_bc5.py -h
usage: test_bc5.py [-h] [-i I] [-e E] [-p P] [-rus RUS] [-ros ROS]
[-rss RSS] [-msl MSL] [-cnn1 CNN1] [-cnn2 CNN2]
[-cnn3 CNN3] [-ft FT] [-wns WNS] [-char CHAR]
[-chari CHARI] [-pos POS] [-posi POSI]
[-rel REL] [-reli RELI] [-dir DIR] [-hd HD]
[-a A]
Multi-channel biLSTM-CNN for relation extraction
optional arguments:
-h, --help show this help message and exit
-i I Job identity
-e E Number of epochs
-p P Patience of early stop (0 for ignore early stop)
-rus RUS Random under sampling number
-ros ROS Random over sampling number
-rss RSS Random seed for re-sampler
-msl MSL Trimmed max sentence length
-cnn1 CNN1 Number of CNN region size 1 filters
-cnn2 CNN2 Number of CNN region size 2 filters
-cnn3 CNN3 Number of CNN region size 3 filters
-ft FT Number of output fastText w2v embedding LSTM dimension
-wns WNS Number of output Wordnet superset LSTM dimension
-char CHAR Number of output character embedding LSTM dimension
-chari CHARI Number of input character embedding LSTM dimension
-pos POS Number of output POS tag LSTM dimension
-posi POSI Number of input POS tag LSTM dimension
-rel REL Number of output dependency relation LSTM dimension
-reli RELI Number of input dependency relation LSTM dimension
-dir DIR Number of dependency direction embedding dimension
-hd HD Hidden layer configurations default '128,128'
-a A Alpha ratio default 0.55
All hyper-parameter is set default to the tuned values. You can change every setting of the model or try the default one.
Example 1: test with default setting on BC5 corpus
$ python test_bc5.py
Example 2: test with some configurations on semeval 2010 corpus
$ python test_semeval.py –ft 128 –cnn1 128 –cnn2 256 –hd 128,256
This command means:
- Change number of output fastText embedding LSTM dimension to 128
- Using 2 region sizes of CNN with 128 and 256 filter corresponding.
- Using 2 hidden layers before softmax layer with 128 and 256 nodes.
After train the model, the model is used to predict on the provided testing dataset. The prediction is written at data/output/
with the name answer-$i
with $i
is the –i
parameter on the running command (default is 0
). The answer file can be used to evaluated with provided testing program.
Command to test the answer:
$ ./evaluate/semeval2010_task8_scorer-v1.2.pl <answer> <key>
The key file is stored at evaluate/KEY
based on the golden testing file.
Example:
$ ./evaluate/semeval2010_task8_scorer-v1.2.pl data/output/answer-example evaluate/KEY
The final result is printed at the last lines with format:
Micro-averaged result (excluding Other):
P = x/x = xx.xx% R = x/x = xx.xx% F1 = xx.xx%
MACRO-averaged result (excluding Other):
P = xx.xx% R = xx.xx% F1 = xx.xx%
<<< The official score is (9+1)-way evaluation with directionality taken into account: macro-averaged F1 = xx.xx% >>>
Official score is of macro-averaged with (9+1)-way evaluation with directionality taken into account. With other ways of evaluation and other result, see the above lines.
For more clear insight evaluation, we re-implement the BC5 evaluation that strictly follows BC5 CDR rules. The evaluated results is printed after training and using model.
The result is printed in the format:
Testing model over test set
result: abstract: (p, r, f1) intra (p, r, f1)
result after apply chemical rule 1/2: abstract: (p, r, f1)
The result contains:
- Results of baseline model with abstract level result (as same as the BC5 testing service) and intra level result (the result conduct on the intra-sentence relation, excluding cross-sentence relation).
- Results after apply chemical rule (as described in the paper) at abstract level (as same as the BC5 testing service).