Natural Language Inference Enhanced by Knowledge Graph Embedding (KEIM)

Dependencies

To run it perfectly, you will need (recommend using Anaconda to set up environment):

• Python 3.5 or 3.6
• Tensorflow 1.10.0
• Java >= 1.6

1. Download and preprocess

cd data_process

Download the data and the resources for preprocess:

• SNLI dataset
• MultiNLI dataset
• SciTail dataset
• GloVe embedding (300D)
• Wordnet 3.0
• CoreNLP

python download.py

Preprocess for SNLI dataset

python preprocess_data_snli.py

Preprocess for MultiNLI dataset

python preprocess_data_multinli.py

Preprocess for SciTail dataset

python process_data_scitail.py

2. Train KEIM

Hyper-parameters are set in configure file in ./config/xxx.sample.config

Training process for SNLI dataset

cd src
python Main.py --config_path ../configs/snli.sample.config

Training process for MultiNLI dataset

python Main.py --config_path ../configs/multinli.sample.config

Training process for SciTail dataset

python Main.py --config_path ../configs/scitail.sample.config

The model and results are saved in $model_dir$.

3. Evaluation

cd src
python Evaluation.py --model_prefix your_model --in_path The path to the test file.