# This folder contains:
1. EntailmentVector: implements about entailment vector
2. TextualEntailment: implements about textual entailment

# This implement contains:
1. Direct evaluation of Entailment Vectors
	a. our baseline: word2vec
	b. NN model
	c. ST model 
2. Evaluation on Textual Entailment
	a. our baseline: w2v
	b. w2v + NN model
	c. w2v + ST model

# Prerequisites
	a. Python 2.7
	b. Tensorflow

# Download data
To download the data, please directly use git clone 
$ git clone https://github.com/lanyunshi/embedding-for-textual-entailment.git
Since the embeddings of SNLI w2vnn and w2vour are too large, please download files w2vnn.txt and w2vour.txt from [link](https://drive.google.com/drive/folders/15XmW6yKVNPJRPppNB_zqR8XsCbehhATT?usp=sharing) and put them under path ./TextualEntailment/data/SNLIglove/

# Usage 
To test pre-trained NN model of Entailment Vectors on test/CS data
$ cd EntailmentVector
$ python code/TrainRunner.py data/ NN only-test -tp test -pm 2017_01_24_10_29_53
change "-tp test" to "-tp CS" is you want to test on CS data

To test pre-trained ST model of Entailment Vectors on test/CS data
$ python code/TrainRunner.py data/ My only-test -tp test -pm 2017_03_21_10_31_00

If you want to train model by yourself
1. convert word pairs data into index
	$ python code/Entity2Idx.py data
2. train word2vec/NN/ST models
	$ python code/TrainRunner.py data/ ConW2v train-test
	$ python code/TrainRunner.py data/ NN train-test
	$ python code/TrainRunner.py data/ My train-test
	# You can change the parameters of model in file "TrainRunner.py". Please set proper parameters to train different models.

To test pre-trained w2v/w2v + NN/w2v + ST model of Textual Entailment on SICK(ss) data
$ cd TextualEntailment
$ python code/train.py data/SICKglove/w2v.txt w2vw2v data/SICK/sequence/train.txt data/SICK/sequence/dev.txt data/SICK/sequence/test.txt saved-model --is_train 0 --load saved-model/sick_w2v
$ python code/train.py data/SICKglove/w2vnn.txt w2vw2vnn data/SICK/sequence/train.txt data/SICK/sequence/dev.txt data/SICK/sequence/test.txt saved-model --is_train 0 --load saved-model/sick_w2vnn
$ python code/train.py data/SICKglove/w2vour.txt w2vw2vour data/SICK/sequence/train.txt data/SICK/sequence/dev.txt data/SICK/sequence/test.txt saved-model --is_train 0 --load saved-model/sick_w2vour

To test pre-trained w2v/w2v + NN/w2v + ST model of Textual Entailment on SICK(ns) data
$ python code/train.py data/SICKglove/w2v.txt w2vw2v data/SICK/sequenceSplit/train.txt data/SICK/sequenceSplit/dev.txt data/SICK/sequenceSplit/test.txt saved-model --is_train 0 --load saved-model/sick_split_w2v
$ python code/train.py data/SICKglove/w2vnn.txt w2vw2vnn data/SICK/sequenceSplit/train.txt data/SICK/sequenceSplit/dev.txt data/SICK/sequenceSplit/test.txt saved-model --is_train 0 --load saved-model/sick_split_w2vnn
$ python code/train.py data/SICKglove/w2vour.txt w2vw2vour data/SICK/sequenceSplit/train.txt data/SICK/sequenceSplit/dev.txt data/SICK/sequenceSplit/test.txt saved-model --is_train 0 --load saved-model/sick_split_w2vour

To test pre-trained w2v/w2v + NN/w2v + ST model of Textual Entailment on SNLI(ss) data
$ python code/train.py data/SNLIglove/w2v.txt w2vw2v data/SNLI/sequence/train.txt data/SNLI/sequence/dev.txt data/SNLI/sequence/test.txt saved-model --is_train 0 --load saved-model/snli_w2v
$ python code/train.py data/SNLIglove/w2vnn.txt w2vw2vnn data/SNLI/sequence/train.txt data/SNLI/sequence/dev.txt data/SNLI/sequence/test.txt saved-model --is_train 0 --load saved-model/snli_w2vnn
$ python code/train.py data/SNLIglove/w2vour.txt w2vw2vour data/SNLI/sequence/train.txt data/SNLI/sequence/dev.txt data/SNLI/sequence/test.txt saved-model --is_train 0 --load saved-model/snli_w2vour

To test pre-trained w2v/w2v + NN/w2v + ST model of Textual Entailment on SNLI(ns) data
$ python code/train.py data/SNLIglove/w2v.txt w2vw2v data/SNLI/sequenceSplit/train.txt data/SNLI/sequenceSplit/dev.txt data/SNLI/sequenceSplit/test.txt saved-model --is_train 0 --load saved-model/snli_split_w2v
$ python code/train.py data/SNLIglove/w2vnn.txt w2vw2vnn data/SNLI/sequenceSplit/train.txt data/SNLI/sequenceSplit/dev.txt data/SNLI/sequenceSplit/test.txt saved-model --is_train 0 --load saved-model/snli_split_w2vnn
$ python code/train.py data/SNLIglove/w2vour.txt w2vw2vour data/SNLI/sequenceSplit/train.txt data/SNLI/sequenceSplit/dev.txt data/SNLI/sequenceSplit/test.txt saved-model --is_train 0 --load saved-model/snli_split_w2vour

To train your own model, you can simply run
$ python code/train.py path/to/golve/file w2vw2v/w2vnn/w2vour path/to/train/data path/to/dev/data path/to/test/data path/to/save-model/folder

# Thank [Erick Rocha Fonseca](https://github.com/erickrf/multiffn-nli)'s implementation of decomposable attention model (DAM), we modified the DAM model based on his original code.