Similarity measures from Babylon Health.

$ pip install simba

You can also checkout this repository and install from the root folder:

$ pip install .

Many of the similarity measures in simba rely on pre-trained embeddings. If you don't have your own encoding logic already, you can register your embedding files to use them easily with simba, as long as they're in the standard text format for word vectors (as described here). For example, if you want to use fastText vectors that you've saved to /path/to/fasttext, you can just run

$ simba embs register --name fasttext --path /path/to/fasttext

and simba will recognise them under the name fasttext.

You can do something similar for frequencies files (like these):

$ simba freqs register --name wiki --path /path/to/wiki/counts

from simba.similarities import dynamax_jaccard
from simba.core import embed

sentences = ('The king has returned', 'Change is good')

# Assuming you've registered fasttext embeddings as described above
x, y = embed([s.split() for s in sentences], embedding='fasttext')
sim = dynamax_jaccard(x, y)

There are more examples, including comparing different similarity metrics on a dataset of pairs, in the examples directory.

This library contains implementations of the following methods in simba.similarities. Please consider citing the corresponding papers in your work if you find them useful.

Papers:

1. Arora et al., ICLR 2017. A Simple but Tough-to-Beat Baseline for Sentence Embeddings
2. Zhelezniak et al., ICLR 2019. Don't Settle for Average, Go for the Max: Fuzzy Sets and Max-Pooled Word Vectors
3. Vargas et al., ICML 2019. Model Comparison for Semantic Grouping
4. Zhelezniak et al., NAACL-HLT 2019. Correlation Coefficients and Semantic Textual Similarity
5. Zhelezniak et al., EMNLP-IJCNLP 2019. Correlations between Word Vector Sets

• April Shen
• Sasho Savkov
• Vitalii Zhelezniak