HoboNet - Relation Classification

Convolution neural network for classification of semantic relationship between two given entities

Test accuracy: 80.06%

Macro F1 score (excluding Other): 84.18%

Step 1:

Prepare train.txt and text.txt run preprocess.py:

python preprocess.py

Step 2:

To get the vector representation of all the words, and then save them as a pickle.

You will need the embeddings we used: https://www.cs.york.ac.uk/nlp/extvec/wiki_extvec.gz

Download the .gz file and place it in a folder called 'embeddings'

python prepare_data.py

The word embeddings, look-up table, training data and test data get saved into a pickle file in ./pkl

Step 3:

To train the cnn model:

python train.py

All the models are saved in ./model

Step 4:

To view the results of the above step:

python test.py

It will display the accuracy and the F1 score on the test dataset

Step 5:

To use the official SemEval2010-Task8 Scorer:

perl scorer/semeval2010_task8_scorer-v1.2.pl results/HoboNet_result.txt scorer/test_key.txt > scorer/result_scores.txt

This will create result_scores.txt file that contains the official scores and confusion matrix for this task

Dependencies:

NumPy
Keras
cPickle
gzip

Use python2

References:

• Pengda Qin, Weiran Xu, Jun Guo, An empirical convolutional neural network approach for semantic relation classification, Neurocomputing, Volume 190, 2016, Pages 1-9, ISSN 0925-2312, http://dx.doi.org/10.1016/j.neucom.2015.12.091
• Ngoc Thang Vu, Heike Adel, Pankaj Gupta, Hinrich Schütze, Combining Recurrent and Convolutional Neural Networks for Relation Classification, arXiv:1605.07333 [cs.CL]
• Cicero Nogueira dos Santos, Bing Xiang, Bowen Zhou, Classifying Relations by Ranking with Convolutional Neural Networks, arXiv:1504.06580 [cs.CL]