salto

Playing with embedding vectors

Installation

You can install this library from PyPI

pip install salto

or from the GitHub repo:

pip install git+https://github.com/krzjoa/salto.git

Motivation

The goal of the salto package is to explore embeddings and check, how the distance between two points (vectors) can be interpreted. We get two arbitrary selected points, such as embedding vectors for ice and fire draw a straight line passing trough the both these points. Then, we treat the newly created line as a new axis by projecting the rest of the points onto this line.

Usage

import numpy as np
import spacy
import salto

nlp = spacy.load('en_core_web_md')

fire = nlp('fire')
ice = nlp('ice')

ice_fire_axis = salto.axis(ice.vector, fire.vector)

cold = ['ice cream', 'polar', 'snow', 'winter', 'fridge', 'Antarctica']
warm = ['boiling water', 'tropical', 'sun', 'summer', 'oven', 'Africa']

cold_vecs = [nlp(w).vector for w in cold]
warm_vecs = [nlp(w).vector for w in warm]

cold_values = [ice_fire_axis(p) for p in cold_vecs]
warm_values = [ice_fire_axis(p) for p in warm_vecs]

ice_fire_axis.plot(
        {'values': cold_values, 'labels': cold, 'color': 'tab:blue'},
        {'values': warm_values, 'labels': warm, 'color': 'tab:red'},
        poles = {'negative': {'label': 'ice', 'color': 'blue'}, 
                 'positive': {'label': 'fire', 'color': 'red'}}
    )