Screen.Recording.2021-06-09.at.4.51.39.mov

clone git@github.com:arcoyk/word2vec.js.git
// open index.html for demo

word2vec.js

Light weight frontend Word2Vec with sample trained model of 3171 Japanese words. 3171 words were randomly selected from original Word2Vec model (330,000 words, 50 dimentional vectors, trained with Wikipedia articles).

• Word count: 3171
• Vector dimention: 50
• Language: Japanese
• Original source: Wikipedia

Pretrained model was translated into words and vectors which can be found at ./js/data50.js. This will be ready as soon as you open index.html. Edit this file to change the model.

Example

// most_similar equivalent
word2vec.w2ws("両極性").ws
// => ["両極端", "メタ理論", "人生観", "欲張り", "主体性"] 

Note that the output is an Object that consists similar words and similarities.

let result = word2vec.w2ws("補修")
// result = {
//     ws: ["補修", "点検", "堰堤", "バラストタンク", "人手"],
//     ps: [1.0000001192092896, 0.7502378225326538, 0.6257367134094238, 0.6090835928916931, 0.5916973352432251]
// }
// similarity between "補修" and "点検" is 0.75023

Offcourse, you can get the vector

word2vec.w2v("補修")
// [-0.14429277181625366, 0.18658098578453064, 0.1847711056470871, 0.09209378808736801, -0.0679391548037529, 0.16327759623527527, -0.031874626874923706, -0.16484060883522034, -0.06378225237131119, 0.023497363552451134, 0.0020464088302105665, 0.17774838209152222, 0.02699054591357708, 0.16259080171585083, 0.28372877836227417, 0.07724079489707947, -0.2403295338153839, -0.0819578692317009, 0.20890462398529053, -0.028673263266682625, 0.009770154021680355, 0.13413363695144653, 0.011923604644834995, -0.11731985211372375, -0.10551074147224426, 0.08331461250782013, 0.004097146913409233, 0.24513141810894012, -0.030268225818872452, -0.01968240551650524, 0.01812598668038845, 0.06748735904693604, 0.09437521547079086, -0.0255605336278677, 0.01223302073776722, -0.016752682626247406, -0.0056835669092834, 0.008190684020519257, -0.4212040603160858, -0.0371280238032341, 0.1373198926448822, -0.1416447013616562, -0.12562932074069977, 0.3315233290195465, 0.12790662050247192, -0.11261098831892014, -0.011190776713192463, -0.17273341119289398, 0.15604254603385925, 0.17403298616409302]

Why

ml5.js provides more stable word2vec on js with full features. word2vec.js was made for ready-to-use repo with a hostable-size model for my personal convenience.

word2vec.js differs...

1. word2vec.js accepts separated model: easy to create tensor matrix with TensorFlow.js, easy to set attribute on WebGL shader (in case you want to visualize it).

// common word2vec model
{"dog": [0.32, 0.43, 0.45], "cat": [0.12, 0.23, 0.54]}
// word2vec.js accepts separated model
{"ws": ["dog", "cat"], "vs": [[0.32, 0.43, 0.45], [0.12, 0.23, 0.54]]}

2. Calculate cosine similarity by matrix dot (which is clean and slightly fast (maybe)).

// Calculation of cosine similarity (for all vectors) can be boiled down to a simple matrix dot which will be done by TensorFlow
word2vec.w2ws = function(w, topn=5) {
    let v = word2vec.w2v(w)
    v = tf.tensor(v)
    v = v.reshape([v.shape[0], 1])
    let p = tf.matMul(Data.tensor.vs, v)
    p = p.flatten()
    let ids = tf.topk(p, topn).indices
    let ws = ids.arraySync().map(i => Data.ws[i])
    let ps = ids.arraySync().map(i => p.arraySync()[i])
    return {ws:ws, ps:ps}
}