Posit AI Weblog: Phrase Embeddings with Keras


Phrase embedding is a technique used to map phrases of a vocabulary to
dense vectors of actual numbers the place semantically related phrases are mapped to
close by factors. Representing phrases on this vector area assist
algorithms obtain higher efficiency in pure language
processing duties like syntactic parsing and sentiment evaluation by grouping
related phrases. For instance, we anticipate that within the embedding area
“cats” and “canine” are mapped to close by factors since they’re
each animals, mammals, pets, and so forth.

On this tutorial we are going to implement the skip-gram mannequin created by Mikolov et al in R utilizing the keras package deal.
The skip-gram mannequin is a taste of word2vec, a category of
computationally-efficient predictive fashions for studying phrase
embeddings from uncooked textual content. We gained’t deal with theoretical particulars about embeddings and
the skip-gram mannequin. If you wish to get extra particulars you may learn the paper
linked above. The TensorFlow Vector Representation of Words tutorial consists of extra particulars as does the Deep Studying With R notebook about embeddings.

There are different methods to create vector representations of phrases. For instance,
GloVe Embeddings are carried out within the text2vec package deal by Dmitriy Selivanov.
There’s additionally a tidy strategy described in Julia Silge’s weblog submit Word Vectors with Tidy Data Principles.

Getting the Knowledge

We’ll use the Amazon Fine Foods Reviews dataset.
This dataset consists of critiques of wonderful meals from Amazon. The information span a interval of greater than 10 years, together with all ~500,000 critiques as much as October 2012. Critiques embrace product and person data, rankings, and narrative textual content.

Knowledge might be downloaded (~116MB) by working:

download.file("https://snap.stanford.edu/information/finefoods.txt.gz", "finefoods.txt.gz")

We’ll now load the plain textual content critiques into R.

Let’s check out some critiques we now have within the dataset.

[1] "I've purchased a number of of the Vitality canned pet food merchandise ...
[2] "Product arrived labeled as Jumbo Salted Peanuts...the peanuts ... 


We’ll start with some textual content pre-processing utilizing a keras text_tokenizer(). The tokenizer shall be
answerable for remodeling every evaluate right into a sequence of integer tokens (which is able to subsequently be used as
enter into the skip-gram mannequin).

tokenizer <- text_tokenizer(num_words = 20000)
tokenizer %>% fit_text_tokenizer(critiques)

Be aware that the tokenizer object is modified in place by the decision to fit_text_tokenizer().
An integer token shall be assigned for every of the 20,000 commonest phrases (the opposite phrases will
be assigned to token 0).

Skip-Gram Mannequin

Within the skip-gram mannequin we are going to use every phrase as enter to a log-linear classifier
with a projection layer, then predict phrases inside a sure vary earlier than and after
this phrase. It might be very computationally costly to output a chance
distribution over all of the vocabulary for every goal phrase we enter into the mannequin. As a substitute,
we’re going to use damaging sampling, that means we are going to pattern some phrases that don’t
seem within the context and prepare a binary classifier to foretell if the context phrase we
handed is actually from the context or not.

In additional sensible phrases, for the skip-gram mannequin we are going to enter a 1d integer vector of
the goal phrase tokens and a 1d integer vector of sampled context phrase tokens. We’ll
generate a prediction of 1 if the sampled phrase actually appeared within the context and 0 if it didn’t.

We’ll now outline a generator operate to yield batches for mannequin coaching.

skipgrams_generator <- operate(textual content, tokenizer, window_size, negative_samples) {
  gen <- texts_to_sequences_generator(tokenizer, sample(textual content))
  operate() {
    skip <- generator_next(gen) %>%
        vocabulary_size = tokenizer$num_words, 
        window_size = window_size, 
        negative_samples = 1
    x <- transpose(skip${couples}) %>% map(. %>% unlist %>% as.matrix(ncol = 1))
    y <- skip$labels %>% as.matrix(ncol = 1)
    list(x, y)

A generator function
is a operate that returns a distinct worth every time it’s known as (generator features are sometimes used to supply streaming or dynamic information for coaching fashions). Our generator operate will obtain a vector of texts,
a tokenizer and the arguments for the skip-gram (the dimensions of the window round every
goal phrase we look at and what number of damaging samples we would like
to pattern for every goal phrase).

Now let’s begin defining the keras mannequin. We’ll use the Keras functional API.

embedding_size <- 128  # Dimension of the embedding vector.
skip_window <- 5       # What number of phrases to think about left and proper.
num_sampled <- 1       # Variety of damaging examples to pattern for every phrase.

We’ll first write placeholders for the inputs utilizing the layer_input operate.

input_target <- layer_input(form = 1)
input_context <- layer_input(form = 1)

Now let’s outline the embedding matrix. The embedding is a matrix with dimensions
(vocabulary, embedding_size) that acts as lookup desk for the phrase vectors.

embedding <- layer_embedding(
  input_dim = tokenizer$num_words + 1, 
  output_dim = embedding_size, 
  input_length = 1, 
  title = "embedding"

target_vector <- input_target %>% 
  embedding() %>% 

context_vector <- input_context %>%
  embedding() %>%

The following step is to outline how the target_vector shall be associated to the context_vector
with the intention to make our community output 1 when the context phrase actually appeared within the
context and 0 in any other case. We wish target_vector to be related to the context_vector
in the event that they appeared in the identical context. A typical measure of similarity is the cosine
. Give two vectors (A) and (B)
the cosine similarity is outlined by the Euclidean Dot product of (A) and (B) normalized by their
magnitude. As we don’t want the similarity to be normalized contained in the community, we are going to solely calculate
the dot product after which output a dense layer with sigmoid activation.

dot_product <- layer_dot(list(target_vector, context_vector), axes = 1)
output <- layer_dense(dot_product, models = 1, activation = "sigmoid")

Now we are going to create the mannequin and compile it.

mannequin <- keras_model(list(input_target, input_context), output)
mannequin %>% compile(loss = "binary_crossentropy", optimizer = "adam")

We are able to see the complete definition of the mannequin by calling abstract:

Layer (kind)                 Output Form       Param #    Linked to                  
input_1 (InputLayer)         (None, 1)          0                                        
input_2 (InputLayer)         (None, 1)          0                                        
embedding (Embedding)        (None, 1, 128)     2560128    input_1[0][0]                 
flatten_1 (Flatten)          (None, 128)        0          embedding[0][0]               
flatten_2 (Flatten)          (None, 128)        0          embedding[1][0]               
dot_1 (Dot)                  (None, 1)          0          flatten_1[0][0]               
dense_1 (Dense)              (None, 1)          2          dot_1[0][0]                   
Whole params: 2,560,130
Trainable params: 2,560,130
Non-trainable params: 0

Mannequin Coaching

We’ll match the mannequin utilizing the fit_generator() operate We have to specify the variety of
coaching steps in addition to variety of epochs we wish to prepare. We’ll prepare for
100,000 steps for five epochs. That is fairly sluggish (~1000 seconds per epoch on a contemporary GPU). Be aware that you just
may additionally get affordable outcomes with only one epoch of coaching.

mannequin %>%
    skipgrams_generator(critiques, tokenizer, skip_window, negative_samples), 
    steps_per_epoch = 100000, epochs = 5
Epoch 1/1
100000/100000 [==============================] - 1092s - loss: 0.3749      
Epoch 2/5
100000/100000 [==============================] - 1094s - loss: 0.3548     
Epoch 3/5
100000/100000 [==============================] - 1053s - loss: 0.3630     
Epoch 4/5
100000/100000 [==============================] - 1020s - loss: 0.3737     
Epoch 5/5
100000/100000 [==============================] - 1017s - loss: 0.3823 

We are able to now extract the embeddings matrix from the mannequin by utilizing the get_weights()
operate. We additionally added row.names to our embedding matrix so we are able to simply discover
the place every phrase is.

Understanding the Embeddings

We are able to now discover phrases which are shut to one another within the embedding. We’ll
use the cosine similarity, since that is what we skilled the mannequin to


find_similar_words <- operate(phrase, embedding_matrix, n = 5) {
  similarities <- embedding_matrix[word, , drop = FALSE] %>%
    sim2(embedding_matrix, y = ., methodology = "cosine")
  similarities[,1] %>% sort(lowering = TRUE) %>% head(n)
find_similar_words("2", embedding_matrix)
        2         4         3       two         6 
1.0000000 0.9830254 0.9777042 0.9765668 0.9722549 
find_similar_words("little", embedding_matrix)
   little       bit       few     small     deal with 
1.0000000 0.9501037 0.9478287 0.9309829 0.9286966 
find_similar_words("scrumptious", embedding_matrix)
scrumptious     tasty fantastic   wonderful     yummy 
1.0000000 0.9632145 0.9619508 0.9617954 0.9529505 
find_similar_words("cats", embedding_matrix)
     cats      canine      children       cat       canine 
1.0000000 0.9844937 0.9743756 0.9676026 0.9624494 

The t-SNE algorithm can be utilized to visualise the embeddings. Due to time constraints we
will solely use it with the primary 500 phrases. To grasp extra in regards to the t-SNE methodology see the article How to Use t-SNE Effectively.

This plot could appear like a multitude, however if you happen to zoom into the small teams you find yourself seeing some good patterns.
Strive, for instance, to discover a group of internet associated phrases like http, href, and so forth. One other group
which may be straightforward to pick is the pronouns group: she, he, her, and so forth.

Leave a Reply

Your email address will not be published. Required fields are marked *