SeqMatchSeq in Tensorflow

This is a tensorflow implementation of SeqMatchSeq model in Learning Natural Language Inference with LSTM.

Environments

• Python 3.x
• TensorFlow 1.2.x

Pre-trained Word Vectors

I use glove.6B shared by Jeffrey Pennington et al.. It can be found at link.

Data

The data used is Stanford Natural Language Inference (SNLI) corpus which can be downloaded at link.

Usage

abandoning useless word vectors

$ python customize_embedding.py --data_dir DATA_DIR --embedding_path EMBEDDING_PATH

training

$ python natural_language_inference.py --ARG=VALUE

evaluating

$ python natural_language_inference.py --forward_only=True --ARG=VALUE

visualizing

$ tensorboard --logdir=DIR

Results

I achieved 81.7415% correct rate on dev set (~3 epochs).