Elang is an acronym that combines the phrases Embedding (E) and Language (Lang) Models. Its goal is to help NLP (natural language processing) researchers, Word2Vec practitioners and data scientists be more productive in training language models. By the 0.1 release (current version, 0.0.8), the package will include ("marked" checkbox indicates a completed feature):

• Visualizing Word2Vec models
  • 2D plot with emphasis on words of interest
  • 2d plot with neighbors of words
  • More coming soon
• Text processing utility
  • Remove stopwords (Indonesian)
  • Remove region entity (Indonesian)
  • Remove calendar words (Indonesian)
  • Remove vulgarity (Indonesian)
• Corpus-building utility
  • Build Indonesian corpus using wikipedia
  • Pre-trained models for quick experimentation

Elang also means "eagle" in Bahasa Indonesia, and the elang Jawa (Javan hawk-eagle) is the national bird of Indonesia, more commonly referred to as Garuda.

The package provides a collection of utility functions and tools that interface with gensim, matplotlib and scikit-learn, as well as curated negative lists for Bahasa Indonesia (kata kasar / vulgar words, stopwords etc) and useful preprocesisng functions. It abstracts away the mundane task so you can train your Word2Vec model faster, and obtain visual feedback on your model more quickly.

Install the latest version of elang:

pip install --upgrade elang

Performing word embeddings in 2 lines of code gets you a visualization:

from elang.plot.utils import plot2d
from gensim.models import Word2Vec

model = Word2Vec.load("path.to.model")
plot2d(model)
# output:

It even looks like a soaring eagle with its outstretched wings!

elang also includes visualization methods to help you visualize a user-defined k number of neighbors to each words. When draggable is set to True, you will obtain a legend that you can move around in the resulting plot.

words = ['bca', 'hitam', 'hutan', 'pisang', 'mobil', "cinta", "pejabat", "android", "kompas"]

plotNeighbours(model, 
    words, 
    method="TSNE", 
    k=15,
    draggable=True)

The plot above plots the 15 nearest neighbors for each word in the supplied words argument. It then renders the plot with a draggable legend.

Because the dimensionality reduction procedure is handled by the underlying sklearn code, you can use any of the valid parameters in the function call to plot2d and plotNeighbours and they will be handed off to the underlying method. Common examples are the perplexity, n_iter and random_state parameters:

model = Word2Vec.load("path.to.model")
bca = model.wv.most_similar("bca", topn=14)
similar_bca = [w[0] for w in bca]
plot2d(
    model,
    method="PCA",
    targets=similar_bca,
    perplexity=20,
    early_exaggeration=50,
    n_iter=2000,
    random_state=0,
)

Output: