Overview

Exposes the gensim Word2Vec model via a RESTful interface. Also caches responses. This only exposes the most_similar method, and supports the positive, negative, and ntop requests.

Usage

Note that startup can take several minutes, as the Word2Vec model needs to be loaded into memory. The server exposes a /keepalive method intended for ELB or other monitoring, and the keepalive will trigger warmup of the Word2Vec model. Peak memory usage is around 7GB.

$ docker run -d -p 8080:80 reutersmedia/pyword2vec:latest

$ curl http://localhost:8080/most_similar?positive=hello&topn=5
{
  "most_similar": [
    [
      "hi", 
      0.6548985838890076
    ], 
    [
      "goodbye", 
      0.6399056911468506
    ], 
    [
      "howdy", 
      0.6310961246490479
    ], 
    [
      "goodnight", 
      0.5920579433441162
    ], 
    [
      "greeting", 
      0.5855876803398132
    ]
  ], 
  "negative_unused": [], 
  "positive_unused": []
}

To query for multiple terms:

$ curl "http://localhost:8080/most_similar?positive=hello&positive=world&topn=5"
{
  "most_similar": [
    [
      "g'day", 
      0.5079617500305176
    ], 
    [
      "globe", 
      0.49662062525749207
    ], 
    [
      "OBAMA_Hello", 
      0.4954448640346527
    ], 
    [
      "hi", 
      0.4945191442966461
    ], 
    [
      "guten_tag", 
      0.4936927855014801
    ]
  ], 
  "negative_unused": [], 
  "positive_unused": []
}

If a term is not present in the Word2Vec model, it will be excluded from the response and the exclusion will be noted in either the "positive_unused" or "negative_unused" list.

$ curl "http://localhost:8080/most_similar?positive=hello&positive=world&positive=flibertygibbet&topn=5"
{
  "most_similar": [
    [
      "g'day", 
      0.5079617500305176
    ], 
    [
      "globe", 
      0.49662062525749207
    ], 
    [
      "OBAMA_Hello", 
      0.4954448640346527
    ], 
    [
      "hi", 
      0.4945191442966461
    ], 
    [
      "guten_tag", 
      0.4936927855014801
    ]
  ], 
  "negative_unused": [], 
  "positive_unused": [
    "flibertygibbet"
  ]
}

Building

In order to build a Docker image, you should download the Google pretrained model to GoogleNews-vectors-negative300.bin.gz and place it in the root directory.

Copyright

Copyright (C) 2017 Thomson Reuters