How words2map derives out-of-vocabulary (OOV) vectors by searching online:
(1) Connect NLP vector database with a web search engine API like Google / Bing
(2) Do a web search on unknown words (just like a human would)
(3) Parse N-grams (e.g. N = 5) for all text from top M websites (e.g. M = 50)
(4) Filter known N-grams from pre-trained corpus (e.g. word2vec, with 3 million N-grams)
(5) Rank N-grams: inverse global frequency x local frequency on M websites (i.e. TF-IDF)
(6) Derive a new vector: sum vectors for top O known N-grams (e.g. O = 25), i.e.
(7) Visualize by reducing dimensions to 2D/3D (e.g t-SNE works, but UMAP recommended)
(8) Finally, show clusters with HDBSCAN, color-coded in a perceptually uniform space
These OOV vectors were derived in a few seconds as explained above:
See this archived blog post for more details on the words2map algorithm.
Derive new vectors for words by searching online
from words2map import *
model = load_model()
words = load_words("passions.csv")
vectors = [derive_vector(word, model) for word in words]
save_derived_vectors(words, vectors, "passions.txt")Analyze derived word vectors
from words2map import *
from pprint import pprint
model = load_derived_vectors("passions.txt")
pprint(k_nearest_neighbors(model=model, k=10, word="Data_Scientists"))Visualize clusters of vectors
from words2map import *
model = load_derived_vectors("passions.txt")
words = [word for word in model.vocab]
vectors = [model[word] for word in words]
vectors_in_2D = reduce_dimensionality(vectors)
generate_clusters(words, vectors_in_2D)Install
# known broken dependencies: automatic conda installation, python 2 -> 3, gensim
# feel free to debug and make a pull request if desired
git clone https://github.com/overlap-ai/words2map.git
cd words2map
./install.sh