How to find similar words in Keras Word Embedding layer

From Stanford's CS244N course, I know Gensim provides a fantastic method to play around the embedding data: most_similar

I was trying to find some equivalent in Keras Embedding layer but I couldn't. It isn't possible out of the box from Keras? Or was it any wrapper on top of it?

Solution

A simple implementation would be:

def most_similar(emb_layer, pos_word_idxs, neg_word_idxs=[], top_n=10):
    weights = emb_layer.weights[0]

    mean = []
    for idx in pos_word_idxs:
        mean.append(weights.value()[idx, :])

    for idx in neg_word_idxs:
        mean.append(weights.value()[idx, :] * -1)

    mean = tf.reduce_mean(mean, 0)

    dists = tf.tensordot(weights, mean, 1)
    best = tf.math.top_k(dists, top_n)

    # Mask words used as pos or neg
    mask = []
    for v in set(pos_word_idxs + neg_word_idxs):
        mask.append(tf.cast(tf.equal(best.indices, v), tf.int8))
    mask = tf.less(tf.reduce_sum(mask, 0), 1)

    return tf.boolean_mask(best.indices, mask), tf.boolean_mask(best.values, mask)

Of course you need to know the indices of the words. I assume you have a word2idx mapping, so you can get them like this: [word2idx[w] for w in pos_words].

To use it:

# Assuming the first layer is the Embedding and you are interested in word with idx 10
idxs, vals = most_similar(model.layers[0], [10])

with tf.Session() as sess:
    init = tf.global_variables_initializer()
    sess.run(init)
    idxs = sess.run(idxs)
    vals = sess.run(vals)

Some potential improvements for that function:

Make sure it returns top_n words (after the mask it returns less words)
gensim uses normalised embeddings (L2_norm)