Pointwise Hilbert–Schmidt Independence Criterion (PHSIC)

PHSIC computes co-occurence strength between sentences with short learning time.

For example, given consistent sentence pairs:

X	Y
They had breakfast at the hotel.	They are full now.
They had breakfast at ten.	I'm full.
She had breakfast with her friends.	She felt happy.
They had breakfast with their friends at the Japanese restaurant.	They felt happy.
He have trouble with his homework.	He cries.
I have trouble associating with others.	I cry.

PHSIC can give high scores to consistent pairs in terms of the given pairs:

X	Y	score
They had breakfast at the hotel.	They are full now.	0.1134
They had breakfast at an Italian restaurant.	They are stuffed now.	0.0023
I have dinner.	I have dinner again.	0.0023

The details can be found in our paper:

Sho Yokoi, Sosuke Kobayashi, Kenji Fukumizu, Jun Suzuki, Kentaro Inui. Pointwise HSIC: A Linear-Time Kernelized Co-occurrence Norm for Sparse Linguistic Expressions. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp.1763–1775 , 2018.

Installation

$ pip install ./

This will install phsic command to your environment:

$ phsic --help

Basic Usage

Download pre-trained wordvecs (e.g. fasttext):

$ wget https://s3-us-west-1.amazonaws.com/fasttext-vectors/crawl-300d-2M.vec.zip
$ unzip crawl-300d-2M.vec.zip

Prepare dataset:

$ TAB="$(printf '\t')"
$ cat << EOF > train.txt
They had breakfast at the hotel.${TAB}They are full now.
They had breakfast at ten.${TAB}I'm full.
She had breakfast with her friends.${TAB}She felt happy.
They had breakfast with their friends at the Japanese restaurant.${TAB}They felt happy.
He have trouble with his homework.${TAB}He cries.
I have trouble associating with others.${TAB}I cry.
EOF
$ cut -f 1 train.txt > train_X.txt
$ cut -f 2 train.txt > train_Y.txt
$ cat << EOF > test.txt
They had breakfast at the hotel.${TAB}They are full now.
They had breakfast at an Italian restaurant.${TAB}They are stuffed now.
I have dinner.${TAB}I have dinner again.
EOF
$ cut -f 1 test.txt > test_X.txt
$ cut -f 2 test.txt > test_Y.txt

Then, train and predict:

$ phsic train_X.txt train_Y.txt --kernel1 Gaussian 1.0 --encoder1 SumBov FasttextEn --emb1 crawl-300d-2M.vec --kernel2 Gaussian 1.0 --encoder2 SumBov FasttextEn --emb2 crawl-300d-2M.vec --limit_words1 10000 --limit_words2 10000 --dim1 3 --dim2 3 --out_prefix toy --out_dir out --X_test test_X.txt --Y_test test_Y.txt
$ cat toy.Gaussian-1.0-SumBov-FasttextEn.Gaussian-1.0-SumBov-FasttextEn.3.3.phsic
1.134489336180434238e-01
2.320408776101631244e-03
2.321869174772554344e-03

Citation

If you make use of this software we would appreciate it if you would cite the paper:

@InProceedings{D18-1203,
  author = 	"Yokoi, Sho
        and Kobayashi, Sosuke
        and Fukumizu, Kenji
        and Suzuki, Jun
        and Inui, Kentaro",
  title = 	"Pointwise HSIC: A Linear-Time Kernelized Co-occurrence Norm for Sparse Linguistic Expressions",
  booktitle = 	"Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
  year = 	"2018",
  publisher = 	"Association for Computational Linguistics",
  pages = 	"1763--1775",
  location = 	"Brussels, Belgium",
  url = 	"http://aclweb.org/anthology/D18-1203"
}