Shortcut-Stacked Sentence Encoders for Multi-Domain Inference

A pytorch implementation for the paper "Shortcut-Stacked Sentence Encoders for Multi-Domain Inference". https://arxiv.org/pdf/1708.02312.pdf

The Data

The data for this model is the SNLI dataset. https://nlp.stanford.edu/projects/snli/

Download the data zip file from here: https://nlp.stanford.edu/projects/snli/snli_1.0.zip and extract it under ./data directory

Embedding model

This model uses a pretrained embedding vector. Specifically the model uses Glove embedding. You download pretrained word vectors from https://nlp.stanford.edu/projects/glove/. According to the paper they have used glove.840B.300d but you can use a smaller one for reducing the computation.

Put the word embedding file under the directory ./models/glove/.

Dependencies

The only packages used are pytorch and tqdm. Tested on Pytorch 1.3 and python 3.5. Code should work on python 3.5+.

How to run

python3.5 ./main.py

Citations

Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP).