1
$\begingroup$

I am quite new to the topic of word embedding using word2vec and models such as skip-gram.

Our teacher introduced us to this TensorFlow code on word2vec which he ran on Google Colab notebook. He changed the vocabulary_size in Step 2 from 50,000 to 500 and we saw a massive improvement in accuracy/reduction in average loss. Could you please explain what causes this improvement?

Is this because the model has lesser words to deal with; hence better probability of words being similar to the target word?

Additionally, increasing the skip window size in Step 4 of skip-gram tends to slightly worsen the accuracy. Why does this happen?

My hypothesis is that as we increase window size, more training is required as we are trying to match more context words next to the target word. In addition, increasing the skip window also affects the depth of contextual information pertaining to the word.

Is my thought process valid?

Thank you.

"""Basic word2vec example."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import argparse
import collections
import hashlib
import math
import os
import random
import sys
from tempfile import gettempdir
import zipfile

import numpy as np
from six.moves import urllib
from six.moves import xrange  # pylint: disable=redefined-builtin
import tensorflow as tf

from tensorflow.contrib.tensorboard.plugins import projector

data_index = 0


def _hash_file(fpath):
  hasher = hashlib.sha256()
  with open(fpath, 'rb') as fpath_file:
    for chunk in iter(lambda: fpath_file.read(65535), b''):
      hasher.update(chunk)
  return hasher.hexdigest()


def word2vec_basic(log_dir):
  """Example of building, training and visualizing a word2vec model."""
  # Create the directory for TensorBoard variables if there is not.
  if not os.path.exists(log_dir):
    os.makedirs(log_dir)

  # Step 1: Download the data.
  # Note: Source website does not support HTTPS right now.
  url = 'http://mattmahoney.net/dc/'

  # pylint: disable=redefined-outer-name
  def maybe_download(filename, expected_bytes, sha256=None):
    """Download a file if not present, and make sure it's the right size."""
    local_filename = os.path.join(gettempdir(), filename)
    if not os.path.exists(local_filename):
      local_filename, _ = urllib.request.urlretrieve(url + filename,
                                                     local_filename)
    statinfo = os.stat(local_filename)

    if sha256 and _hash_file(local_filename) != sha256:
      raise Exception('Failed to verify ' + local_filename + ' due to hash '
                      'mismatch. Can you get to it with a browser?')

    if statinfo.st_size == expected_bytes:
      print('Found and verified', filename)
    else:
      print(statinfo.st_size)
      raise Exception('Failed to verify ' + local_filename +
                      '. Can you get to it with a browser?')
    return local_filename

  filename = maybe_download(
      'text8.zip',
      31344016,
      sha256='a6640522afe85d1963ad56c05b0ede0a0c000dddc9671758a6cc09b7a38e5232')

  # Read the data into a list of strings.
  def read_data(filename):
    """Extract the first file enclosed in a zip file as a list of words."""
    with zipfile.ZipFile(filename) as f:
      data = tf.compat.as_str(f.read(f.namelist()[0])).split()
    return data

  vocabulary = read_data(filename)
  print('Data size', len(vocabulary))

  # Step 2: Build the dictionary and replace rare words with UNK token.
  vocabulary_size = 50000

  def build_dataset(words, n_words):
    """Process raw inputs into a dataset."""
    count = [['UNK', -1]]
    count.extend(collections.Counter(words).most_common(n_words - 1))
    dictionary = {word: index for index, (word, _) in enumerate(count)}
    data = []
    unk_count = 0
    for word in words:
      index = dictionary.get(word, 0)
      if index == 0:  # dictionary['UNK']
        unk_count += 1
      data.append(index)
    count[0][1] = unk_count
    reversed_dictionary = dict(zip(dictionary.values(), dictionary.keys()))
    return data, count, dictionary, reversed_dictionary

  # Filling 4 global variables:
  # data - list of codes (integers from 0 to vocabulary_size-1).
  #   This is the original text but words are replaced by their codes
  # count - map of words(strings) to count of occurrences
  # dictionary - map of words(strings) to their codes(integers)
  # reverse_dictionary - map of codes(integers) to words(strings)
  data, count, unused_dictionary, reverse_dictionary = build_dataset(
      vocabulary, vocabulary_size)
  del vocabulary  # Hint to reduce memory.
  print('Most common words (+UNK)', count[:5])
  print('Sample data', data[:10], [reverse_dictionary[i] for i in data[:10]])

  # Step 3: Function to generate a training batch for the skip-gram model.
  def generate_batch(batch_size, num_skips, skip_window):
    global data_index
    assert batch_size % num_skips == 0
    assert num_skips <= 2 * skip_window
    batch = np.ndarray(shape=(batch_size), dtype=np.int32)
    labels = np.ndarray(shape=(batch_size, 1), dtype=np.int32)
    span = 2 * skip_window + 1  # [ skip_window target skip_window ]
    buffer = collections.deque(maxlen=span)  # pylint: disable=redefined-builtin
    if data_index + span > len(data):
      data_index = 0
    buffer.extend(data[data_index:data_index + span])
    data_index += span
    for i in range(batch_size // num_skips):
      context_words = [w for w in range(span) if w != skip_window]
      words_to_use = random.sample(context_words, num_skips)
      for j, context_word in enumerate(words_to_use):
        batch[i * num_skips + j] = buffer[skip_window]
        labels[i * num_skips + j, 0] = buffer[context_word]
      if data_index == len(data):
        buffer.extend(data[0:span])
        data_index = span
      else:
        buffer.append(data[data_index])
        data_index += 1
    # Backtrack a little bit to avoid skipping words in the end of a batch
    data_index = (data_index - span) % len(data)
    return batch, labels

  batch, labels = generate_batch(batch_size=8, num_skips=2, skip_window=1)
  for i in range(8):
    print(batch[i], reverse_dictionary[batch[i]], '->', labels[i, 0],
          reverse_dictionary[labels[i, 0]])

  # Step 4: Build and train a skip-gram model.

  batch_size = 128
  embedding_size = 128  # Dimension of the embedding vector.
  skip_window = 1  # How many words to consider left and right.
  num_skips = 2  # How many times to reuse an input to generate a label.
  num_sampled = 64  # Number of negative examples to sample.

  # We pick a random validation set to sample nearest neighbors. Here we limit
  # the validation samples to the words that have a low numeric ID, which by
  # construction are also the most frequent. These 3 variables are used only for
  # displaying model accuracy, they don't affect calculation.
  valid_size = 16  # Random set of words to evaluate similarity on.
  valid_window = 100  # Only pick dev samples in the head of the distribution.
  valid_examples = np.random.choice(valid_window, valid_size, replace=False)

  graph = tf.Graph()

  with graph.as_default():

    # Input data.
    with tf.name_scope('inputs'):
      train_inputs = tf.placeholder(tf.int32, shape=[batch_size])
      train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])
      valid_dataset = tf.constant(valid_examples, dtype=tf.int32)

    # Ops and variables pinned to the CPU because of missing GPU implementation
    with tf.device('/cpu:0'):
      # Look up embeddings for inputs.
      with tf.name_scope('embeddings'):
        embeddings = tf.Variable(
            tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))
        embed = tf.nn.embedding_lookup(embeddings, train_inputs)

      # Construct the variables for the NCE loss
      with tf.name_scope('weights'):
        nce_weights = tf.Variable(
            tf.truncated_normal([vocabulary_size, embedding_size],
                                stddev=1.0 / math.sqrt(embedding_size)))
      with tf.name_scope('biases'):
        nce_biases = tf.Variable(tf.zeros([vocabulary_size]))

    # Compute the average NCE loss for the batch.
    # tf.nce_loss automatically draws a new sample of the negative labels each
    # time we evaluate the loss.
    # Explanation of the meaning of NCE loss and why choosing NCE over tf.nn.sampled_softmax_loss:
    #   http://mccormickml.com/2016/04/19/word2vec-tutorial-the-skip-gram-model/
    #   http://papers.nips.cc/paper/5165-learning-word-embeddings-efficiently-with-noise-contrastive-estimation.pdf
    with tf.name_scope('loss'):
      loss = tf.reduce_mean(
          tf.nn.nce_loss(
              weights=nce_weights,
              biases=nce_biases,
              labels=train_labels,
              inputs=embed,
              num_sampled=num_sampled,
              num_classes=vocabulary_size))

    # Add the loss value as a scalar to summary.
    tf.summary.scalar('loss', loss)

    # Construct the SGD optimizer using a learning rate of 1.0.
    with tf.name_scope('optimizer'):
      optimizer = tf.train.GradientDescentOptimizer(1.0).minimize(loss)

    # Compute the cosine similarity between minibatch examples and all
    # embeddings.
    norm = tf.sqrt(tf.reduce_sum(tf.square(embeddings), 1, keepdims=True))
    normalized_embeddings = embeddings / norm
    valid_embeddings = tf.nn.embedding_lookup(normalized_embeddings,
                                              valid_dataset)
    similarity = tf.matmul(
        valid_embeddings, normalized_embeddings, transpose_b=True)

    # Merge all summaries.
    merged = tf.summary.merge_all()

    # Add variable initializer.
    init = tf.global_variables_initializer()

    # Create a saver.
    saver = tf.train.Saver()

  # Step 5: Begin training.
  num_steps = 100001

  with tf.compat.v1.Session(graph=graph) as session:
    # Open a writer to write summaries.
    writer = tf.summary.FileWriter(log_dir, session.graph)

    # We must initialize all variables before we use them.
    init.run()
    print('Initialized')

    average_loss = 0
    for step in xrange(num_steps):
      batch_inputs, batch_labels = generate_batch(batch_size, num_skips,
                                                  skip_window)
      feed_dict = {train_inputs: batch_inputs, train_labels: batch_labels}

      # Define metadata variable.
      run_metadata = tf.RunMetadata()

      # We perform one update step by evaluating the optimizer op (including it
      # in the list of returned values for session.run()
      # Also, evaluate the merged op to get all summaries from the returned
      # "summary" variable. Feed metadata variable to session for visualizing
      # the graph in TensorBoard.
      _, summary, loss_val = session.run([optimizer, merged, loss],
                                         feed_dict=feed_dict,
                                         run_metadata=run_metadata)
      average_loss += loss_val

      # Add returned summaries to writer in each step.
      writer.add_summary(summary, step)
      # Add metadata to visualize the graph for the last run.
      if step == (num_steps - 1):
        writer.add_run_metadata(run_metadata, 'step%d' % step)

      if step % 2000 == 0:
        if step > 0:
          average_loss /= 2000
        # The average loss is an estimate of the loss over the last 2000
        # batches.
        print('Average loss at step ', step, ': ', average_loss)
        average_loss = 0

      # Note that this is expensive (~20% slowdown if computed every 500 steps)
      if step % 10000 == 0:
        sim = similarity.eval()
        for i in xrange(valid_size):
          valid_word = reverse_dictionary[valid_examples[i]]
          top_k = 8  # number of nearest neighbors
          nearest = (-sim[i, :]).argsort()[1:top_k + 1]
          log_str = 'Nearest to %s:' % valid_word

          print(
              log_str,
              ', '.join([reverse_dictionary[nearest[k]] for k in range(top_k)]))
    final_embeddings = normalized_embeddings.eval()

    # Write corresponding labels for the embeddings.
    with open(log_dir + '/metadata.tsv', 'w') as f:
      for i in xrange(vocabulary_size):
        f.write(reverse_dictionary[i] + '\n')

    # Save the model for checkpoints.
    saver.save(session, os.path.join(log_dir, 'model.ckpt'))

    # Create a configuration for visualizing embeddings with the labels in
    # TensorBoard.
    config = projector.ProjectorConfig()
    embedding_conf = config.embeddings.add()
    embedding_conf.tensor_name = embeddings.name
    embedding_conf.metadata_path = os.path.join(log_dir, 'metadata.tsv')
    projector.visualize_embeddings(writer, config)

  writer.close()

  # Step 6: Visualize the embeddings.

  # pylint: disable=missing-docstring
  # Function to draw visualization of distance between embeddings.
  def plot_with_labels(low_dim_embs, labels, filename):
    assert low_dim_embs.shape[0] >= len(labels), 'More labels than embeddings'
    plt.figure(figsize=(18, 18))  # in inches
    for i, label in enumerate(labels):
      x, y = low_dim_embs[i, :]
      plt.scatter(x, y)
      plt.annotate(
          label,
          xy=(x, y),
          xytext=(5, 2),
          textcoords='offset points',
          ha='right',
          va='bottom')

    plt.savefig(filename)

  try:
    # pylint: disable=g-import-not-at-top
    from sklearn.manifold import TSNE
    import matplotlib.pyplot as plt

    tsne = TSNE(
        perplexity=30, n_components=2, init='pca', n_iter=5000, method='exact')
    plot_only = 500
    low_dim_embs = tsne.fit_transform(final_embeddings[:plot_only, :])
    labels = [reverse_dictionary[i] for i in xrange(plot_only)]
    plot_with_labels(low_dim_embs, labels, os.path.join(gettempdir(),
                                                        'tsne.png'))

  except ImportError as ex:
    print('Please install sklearn, matplotlib, and scipy to show embeddings.')
    print(ex)


# All functionality is run after tf.compat.v1.app.run() (b/122547914). This
# could be split up but the methods are laid sequentially with their usage for
# clarity.
def main(unused_argv):
  # Give a folder path as an argument with '--log_dir' to save
  # TensorBoard summaries. Default is a log folder in current directory.
  current_path = os.path.dirname(os.path.realpath(sys.argv[0]))

  parser = argparse.ArgumentParser()
  parser.add_argument(
      '--log_dir',
      type=str,
      default=os.path.join(current_path, 'log'),
      help='The log directory for TensorBoard summaries.')
  flags, unused_flags = parser.parse_known_args()
  word2vec_basic(flags.log_dir)


if __name__ == '__main__':
  tf.app.run()
$\endgroup$

1 Answer 1

1
$\begingroup$

The less-common words don't carry much information, because they don't appear frequently around other words; they don't appear frequently at all. Also sometimes low-count words are typos that are going to be noise anyway.

Trying to learn about them isn't helping much, and may be prone to overfitting their embeddings because they appear rarely, which hurts generalization.

Treating them as all one big 'unknown' word doesn't lose much, and allows for a simpler network to learn the embedding, which is faster.

If the window is too wide, it's learning positive associations between words further away in the text, but if they're too far away that's probably not a good assumption, and hurts the quality of the model.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.