aliosm/semantic-question-similarity

Official implementation of: Tha3aroon at NSURL-2019 Task 8: Semantic Question Similarity in Arabic

Semantic-Question-Similarity

The official implementation of our paper: Tha3aroon at NSURL-2019 Task 8: Semantic Question Similarity in Arabic, which was part of NSURL-2019 workshop on Task 8 for Arabic Semantic Question Similarity.

0. Prerequisites

• Python >= 3.6
• Install required packages listed in requirements.txt file
  • pip install -r requirements.txt
• To use ELMo embeddings:
  • Clone ELMoForManyLangs repository
    • git clone https://github.com/HIT-SCIR/ELMoForManyLangs.git
  • Install the package:
    • cd ELMoForManyLangs
    • python setup.py install
    • cd ..
  • Download and unzip Arabic pre-trainled ELMo model
    • wget http://vectors.nlpl.eu/repository/11/136.zip -O elmo_dir/136.zip
    • unzip elmo_dir/136.zip -d elmo_dir
    • cp ELMoForManyLangs/configs/cnn_50_100_512_4096_sample.json elmo_dir/cnn_50_100_512_4096_sample.json

1. Data Preprocessing

Data preprocessing step to separate punctuations from words

python 1_preprocess.py --dataset-split train
python 1_preprocess.py --dataset-split test

2. Data Enlarging

Enlarging the data using both Positive and Negative Transitive properties (descriped in the paper)

python 2_enlarge.py

3. Generating Words Embeddings

To make the training step faster, we pre-generate words embeddings from either ELMo or BERT models and store them in a pickle file

python 3_build_embeddings_dict.py --embeddings-type elmo # For ELMo
python 3_build_embeddings_dict.py --embeddings-type bert # For BERT

We adopted using ELMoForManyLangs over bert-embedding because it yields better results.

4. Model Training

Training the model using ELMo with 0.2 dropout, 256 batch size, 100 epochs and 2000 dev set size

python 4_train.py --embeddings-type elmo --dropout-rate 0.2 --batch-size 256 --epochs 100 --dev-split 2000

This hyperparameters setup gives the best results according to our experiments, change the values in order to experiment more..

5. Model Inferencing

Inferencing predictions for the test set is done given the path to a certain checkpoint, the default threshold is 0.5 which can be changed using the optional argument --threshold

python 5_infer.py --model-path checkpoints/epoch100.h5

Model Structure

The following figure illustrates our best model structure.

Note: All codes in this repository are tested on Ubuntu 18.04

Contributors

1. Ali Hamdi Ali Fadel.
   
2. Ibraheem Tuffaha.
   
3. Mahmoud Al-Ayyoub.
   

License

The project is available as open source under the terms of the MIT License.