tensorflow-hub - 如何使用 TF Hub 模块找到最接近的单词？

Question

如果嵌入之间的 np.inner 返回它们的相似性。如何从一个词嵌入中找到最接近的词？我正在使用 Wiki 模块。

Answer 1

您可以执行以下操作：

# Download 10k most popular english words.    
import urllib2
response = urllib2.urlopen('https://raw.githubusercontent.com/first20hours/google-10000-english/master/google-10000-english-no-swears.txt')
words = response.read().split("\n")

# Compute the embedding for 10k words.
with tf.Graph().as_default():
  embed = hub.Module("https://tfhub.dev/google/Wiki-words-250/1")
  embeddings = embed(words)
  with tf.train.MonitoredSession() as sess:
    values = sess.run(embeddings)

data = dict(zip(words, values))

def get_neighbors(all_data, target, count=3):
  # Compute points in all_data that are the closest to "target".
  # Sort the words based on the length of the vector between them.
  # Then pick "count" closest ones.
  return sorted(all_data, key=lambda emb_key: np.linalg.norm(all_data[emb_key] - target))[:count]

# Then make queries of your choice, e.g.
print(get_neighbors(data, data["first"]))
# Prints ['first', 'second', 'third'] 
print(get_neighbors(data, data["woman"] - data["man"] + data["father"], count=10))
# Prints ['mother', 'father', 'wife', 'daughter', 'wed', 'husband', 'uncle', 'son', 'child', 'mistress']