iParaphrasing

scripts for our COLING paper "iParaphrasing: Extracting Visually Grounded Paraphrases via an Image"

Dependencies

This scripts are tested on

• python 3.6
• numpy 1.13.3
• scipy 1.0.0
• scikit-learn 0.19.1
• pandas 0.21.0
• chainer 4.0.0
• GPy 1.9.2
• GPuOpt 1.2.5

Data

Get Flickr30K entities dataset here.

Other materials can be downloaded from here (figshare).

Download data.zip, ari_data.zip and models.zip, then extract zip files under coling_iparaphrasing.

Train a model

Run the command below in coling_iparaphrasing directory．

FlickrIMG_ROOT=/path/to/flickr30k-images/ python codes/script/training/train_paraphrase_classifier.py -d 0 --image_net vgg --phrase_net fv+cca

By default, the output model and log files will be stored under checkpoint/generated_name/

For more details, run

python codes/script/training/train_paraphrase_classifier.py --help

Get prediction

FlickrIMG_ROOT=/path/to/flickr30k-images/ python codes/script/training/train_paraphrase_classifier.py -d 0 --eval /path/to/output/directory/

Prediction results will be written to res_test.csv in the model directory.

Evaluate prediction

First, run

codes/script/shell/prepare_eval.sh

See codes/notebook/[COLING] Table 1.ipynb and codes/notebook/[COLING] Table 1-ARI scores.ipynb