This repository is connected with Chinese Definition Modeling task.

The src directory consists of three models:

• baseline
• Adaptive-Attention Model
• Self- and Adaptive-Attention Model

Paper Link: https://arxiv.org/abs/1905.06512

Contact: [email protected]

• python (3.6)
• xlrd (1.1.0)
• jieba (0.39)
• progressbar2 (3.38.0)

The dataset construction procedure follows the README.mdfile in the scripts/make_dataset directory.

We have also written an integreted script make_dataset.sh in the directory of src.

cd src
chmod +x make_dataset.sh
./make_dataset.sh

The CWN dataset we used in the experiments is in the dataset/cwn directory.

The baseline model is based on Websail-NU/torch-defseq, and detailed instruction can be found there.

The Adaptive-Attention model is in the directory of src/aam, and can run as follows:

• Requirements

  • python (2.7)
  • pytorch (0.3.1)
  • numpy (1.14.5)
  • gensim (3.5.0)
  • kenlm

• Preprocess

  The preprocess procedure is written in the script of preprocess.sh. During preprocessing, we used pretrained Chinese word embeddings, which is trained on the Chinese Gigaword Corpus. Jieba Chinese segmentation tool is employed. The binarized word2vec file is named gigaword_300d_jieba.bin placed in the directory of data.

  cd src/adaptive
  ./preprocess.sh

• Training & Inference

  You can use following commands to train and inference. Also, we've uploaded the training_lot.txt of the best model in the directory of models/adaptive/best.

  ./train.sh best #using the best parameters to train a model
  ./inference.sh best 22 #22 denotes the best epoch

• Scoring

  • A function_words.txt is needed in the data directory, we've extracted one from the HowNet when making the dataset
  • A chinesegigawordv5.lm Chinese language model is needed in the data directory, any arpa format language model will do
  • Then you can use the following script to compute the score of BLEU

  ./score.sh best 21 #21 denotes the best epoch

The Self- and Adaptive-Attention Model is in the directory of src/saam. The instruction of this model is as follows:

• Requirements and Installation

  • python (3.6)
  • pytorch (0.4.1)
  • use following commands to install other requirements

  cd src/self-attention
  pip install -r requirements.txt

• Preprocess

  The preprocessing scripts is used to convert text files into binarized data.

  ./preprocess.sh

• Train & Generate

  We use fixed pre-trained word embeddings as the adaptive attention model. The word embedding is in the directory of data and named chinesegigawordv5.jieba.skipngram.300d.txt. We uploaded a demo word embedding file which contains only 100 lines.

  The model can be trained and employed using following commands:

  ./train.sh best #best is name of the model
  ./generate.sh best

  Parameters used for training is written in the train.sh script