match-lstm

Pytorch Implementation of "Learning Natural Language Inference with LSTM", 2016, S. Wang et al. (https://arxiv.org/pdf/1512.08849.pdf)

Env.

• Python 3

Requirements

• PyTorch 1.2.0
• numpy

Dataset

• Download snli_1.0.zip (90.2 MB) and decompress snli_1.0_train.txt, snli_1.0_dev.txt and snli_1.0_test.txt to data/
  • More information can be found at https://nlp.stanford.edu/projects/snli/

Word Embeddings

• Download glove.840B.300d.zip (2.0 GB) and decompress glove.840B.300d.txt to $HOME/common/
  • See https://nlp.stanford.edu/projects/glove/

Experiment

# Create a pickle file: data/snli.pkl
$ python3 dataset.py

# Run
$ python3 main.py

Training time

• 156 minutes per training epoch w/ a NVIDIA Titan Xp GPU
• I plan to reduce the training time soon. (Work in progress)

Result

• Epoch 6
• Training loss: 0.361281, accuracy: 86.1% (mLSTM train accuracy: 92.0%)
• Dev loss: 0.392275, accuracy: 85.8% (mLSTM dev accuracy: 86.9%)
• Test loss: 0.397926, accuracy: 85.5% (mLSTM test accuracy: 86.1%)

Reference

https://github.com/shuohangwang/SeqMatchSeq