The source code of this tutorial can be found at [book/word2vec](https://github.com/PaddlePaddle/book/tree/develop/word2vec), please refer to PaddlePaddle [installation guide](http://www.paddlepaddle.org/doc_cn/build_and_install/index.html) for the first time use.
This is intended as a reference tutorial. The source code of this tutorial lives on [book/word2vec](https://github.com/PaddlePaddle/book/tree/develop/word2vec).
For instructions on getting started with PaddlePaddle, see [PaddlePaddle installation guide](http://www.paddlepaddle.org/doc_cn/build_and_install/index.html).
## Background Introduction
## Background Introduction
In this chapter, we will introduce the vector representation of words, which is also known as word embedding. Word embedding is a common operation in natural language processing, and is the common fundamental technology behind the search engine, ads system, recommendation system and other Internet services.
This section introduces the concept of **word embedding**, which is a vector representation of words. It is a popular technique used in natural language processing. Word embeddings support many Internet services, including search engines, advertising systems, and recommendation systems.
### One-Hot Vectors
Building these services requires us to quantify the similarity between two words or paragraphs. This calls for a new representation of all the words to make them more suitable for computation. An obvious way to achieve this is through the vector space model, where every word is represented as an **one-hot vector**.
In those Internet services, we usually need to estimate the similarity between two words or two paragraphs. In order to perform this estimation, we need to first represent the words in a way that is easier for the computer to process. The most natural way to achieve this goal is vector space model. In this kind of model, every word is represented as a one-hot vector, whose length is the size of the dictionary. Every element in the one-hot vector corresponds to one word in the dictionary. For each word, the corresponding element in the vector is 1 and all other elements are 0.
For each word, its vector representation has the corresponding entry in the vector as 1, and all other entries as 0. The lengths of one-hot vectors match the size of the dictionary. Each entry of a vector corresponds to the presence (or absence) of a word in the dictionary.
Although one-hot vector is a natural choice, it has limited usefulness. For example, in an Internet ads system, if a customer enters a query of "Mother's day", and the keyword of an ad is "Carnations". We know that these two words are connected based on common sense because normally people would send their mothers a bunch of carnations on mother's day. However, the metric distance (either Euclidean or cosine similarity) between the one-hot vectors of these two words indicates that they are not relevant to each other. The reason that we came to this counter-intuitive conclusion is that the information contained in each word is too small. Therefore, only comparing two words is not sufficient for us to accurately estimate their relevance. In order to accurately calculate their similarity, we need more information which could be learned from big data through machine learning method.
One-hot vectors are intuitive, yet they have limited usefulness. Take the example of an Internet advertising system: Suppose a customer enters the query "Mother's Day", while an ad bids for the keyword carnations". Because the one-hot vectors of these two words are perpendicular, the metric distance (either Euclidean or cosine similarity) between them would indicate little relevance. However, *we* know that these two queries are connected semantically, since people often gift their mothers bundles of carnation flowers on Mother's Day. This discrepancy is due to the low information capacity in each vector. That is, comparing the vector representations of two words does not assess their relevance sufficiently. To calculate their similarity accurately, we need more information, which could be learned from large amounts of data through machine learning methods.
In the machine learning field, different kinds of "knowledge" are represented by different kinds of model, and word embedding model is one of them. Word embedding model can map an one-hot vector to an embedding vector of lower dimension, like $embedding(mother's day) = [0.3, 4.2, -1.5, ...], embedding(carnations) = [0.2, 5.6, -2.3, ...]$. In this mapped embedding vector space, we wish that the embedding vectors of two similar words (in terms of either semantic meaning or usage) are more close to each other, so that the cosine similarity between the corresponding vectors for "mother's day" and "carnations" are not zero anymore.
Like many machine learning models, word embeddings can represent knowledge in various ways. Another model may project an one-hot vector to an embedding vector of lower dimension e.g. $embedding(mother's day) = [0.3, 4.2, -1.5, ...], embedding(carnations) = [0.2, 5.6, -2.3, ...]$. Mapping one-hot vectors onto an embedded vector space has the potential to bring the embedding vectors of similar words (either semantically or usage-wise) closer to each other, so that the cosine similarity between the corresponding vectors for words like "Mother's Day" and "carnations" are no longer zero.
Word embedding model could be probabilistic model, co-occurrence matrix model or neural network model. Before using neural networks to calculate the word embedding, the traditional method is to calculate a co-occurrence matrix $X$ of words. $X$ is a $|V| \times |V|$ size of matrix, where $X_{ij}$ represents the co-occurrence times of the i-th and j-th word in the vocabulary `V` within all corpus, and $|V|$ is the size of the vocabulary. By performing matrix decomposition on $X$ (like Singular Value Decomposition \[[5](#References)\]), the resulting $U$ can be seen as the word embedding of all the words.
A word embedding model could be a probabilistic model, a co-occurrence matrix model, or a neural network. Before people started using neural networks to generate word embedding, the traditional method was to calculate a co-occurrence matrix $X$ of words. Here, $X$ is a $|V| \times |V|$ matrix, where $X_{ij}$ represents the co-occurrence times of the $i$th and $j$th words in the vocabulary `V` within all corpus, and $|V|$ is the size of the vocabulary. By performing matrix decomposition on $X$ e.g. Singular Value Decomposition \[[5](#References)\]
$$X = USV^T$$
$$X = USV^T$$
However such traditional method suffers from many drawbacks:
the resulting $U$ can be seen as the word embedding of all the words.
1) Since lots of words didn't have co-occurrences, the matrix is extremely sparse. So it would require further treatment on word frequency to achieve good performance of matrix factorization;
2) The matrix size is huge (normally on the order of $10^6*10^6$);
3) We need to manually delete stop words (like "although", "a", ...), otherwise these frequent words will affect the performance of matrix factorization.
The neural network based model does not need to store a huge statistic table on all of the corpus. It obtains the word embedding by learning from semantic information, hence could avoid the aforementioned problems in the traditional method. In this chapter, we will introduce the details of neural network word embedding model and how to train such model in PaddlePaddle.
However, this method suffers from many drawbacks:
1) Since many pairs of words don't co-occur, the co-occurrence matrix is sparse. To achieve good performance of matrix factorization, further treatment on word frequency is needed;
2) The matrix is large, frequently on the order of $10^6*10^6$;
3) We need to manually filter out stop words (like "although", "a", ...), otherwise these frequent words will affect the performance of matrix factorization.
The neural network based model does not require storing huge hash tables of statistics on all of the corpus. It obtains the word embedding by learning from semantic information, hence could avoid the aforementioned problems in the traditional method. In this chapter, we will introduce the details of neural network word embedding model and how to train such model in PaddlePaddle.
## Results Demonstration
## Results Demonstration
In this chapter, after training the word embedding model, we could use the data visualization algorithm t-SNE\[[4](#reference)\] to draw the word embedding vectors after projecting them onto a two-dimensional space (see figure below). From the figure we could see that the semantically relevant words (like 'a', 'the', 'these'; 'big', 'huge') are close to each other in the projected space, while irrelevant words (like 'say', 'business'; 'decision', 'japan') are far from each other.
In this section, after training the word embedding model, we could use the data visualization algorithm $t-$SNE\[[4](#reference)\] to draw the word embedding vectors after projecting them onto a two-dimensional space (see figure below). From the figure we could see that the semantically relevant words -- *a*, *the*, and *these* or *big* and *huge* -- are close to each other in the projected space, while irrelevant words -- *say* and *business* or *decision* and *japan* -- are far from each other.
Figure 1. Two dimension projection of word embedding
Figure 1. Two dimension projection of word embeddings
</p>
</p>
On the other hand, we know that the cosine similarity between two vectors falls between $[-1,1]$: the cosine similarity of two identical vectors is 1, of two perpendicular vectors is 0, of two opposite vectors is -1, which means that the cosine similarity between two vectors is proportional to their relevance. So we can calculate the cosine similarity of two word embedding vectors:
### Cosine Similarity
On the other hand, we know that the cosine similarity between two vectors falls between $[-1,1]$. Specifically, the cosine similarity is 1 when the vectors are identical, 0 when the vectors are perpendicular, -1 when the are of opposite directions. That is, the cosine similarity between two vectors scales with their relevance. So we can calculate the cosine similarity of two word embedding vectors to represent their relevance:
```
```
please input two words: big huge
please input two words: big huge
...
@@ -42,83 +52,90 @@ please input two words: from company
...
@@ -42,83 +52,90 @@ please input two words: from company
similarity: -0.0997506977351
similarity: -0.0997506977351
```
```
The above results could be obtained by running `calculate_dis.py` which loads the words in the dictionary and their corresponding trained word embeddings. We will provide detailed instruction in the section of [Model Application](#Model Application)
The above results could be obtained by running `calculate_dis.py`, which loads the words in the dictionary and their corresponding trained word embeddings. For detailed instruction, see section [Model Application](#Model Application).
## Model Overview
## Model Overview
In this section we will introduce three word embedding models: N-gram model, CBOW model and Skip-gram model, which shares a common theme of getting the probability of a word given its context. For N-gram model, we will first introduce the concept of language model, and implement it using PaddlePaddle in the following section of [Model Training](#Model Training). The latter two models are neural word embedding model which became popular recently and was developed by Tomas Mikolov at Google \[[3](#reference)\]. Although their structures are shallow and simple, they are very effective.
In this section, we will introduce three word embedding models: N-gram model, CBOW, and Skip-gram, which all output the frequency of each word given its immediate context.
For N-gram model, we will first introduce the concept of language model, and implement it using PaddlePaddle in section [Model Training](#Model Training).
The latter two models, which became popular recently, are neural word embedding model developed by Tomas Mikolov at Google \[[3](#reference)\]. Despite their apparent simplicity, these models train very well.
### Language Model
### Language Model
Before introducing the word embedding model, we will first introduce a concept: language model. Language model aims at modeling the joint probability function $P(w_1, ..., w_T)$ of a sentence, where $w_i$ is the i-th word in the sentence. The goal of the language model is to give meaningful sentences higher probabilities and meaningless sentences lower probabilities. Such kind of model can be applied to many fields, like machine translation, speech recognition, information retrieval, part-of-speech tagging and handwriting recognition, all of which require the probability of a sequence. Let us take information retrieval for example. When you search "how long is a football bame" (bame is a medical word), the search engine will ask you if you would actually like to search "how long is a football game" instead. This is because the probability of "how long is a football bame" is very low according to the language model, and among all of the words close to "bame", the word "game" would make the probability of the sentence highest.
Before diving into word embedding models, we will first introduce the concept of **language model**. Language models build the joint probability function $P(w_1, ..., w_T)$ of a sentence, where $w_i$ is the i-th word in the sentence. The goal is to give higher probabilities to meaningful sentences, and lower probabilities to meaningless constructions.
For language model's target probability $P(w_1, ..., w_T)$, if we assume that each word in the sentence is independent, then the joint probability of the whole sentence is the product of each word's probability:
In general, models that generate the probability of a sequence can be applied to many fields, like machine translation, speech recognition, information retrieval, part-of-speech tagging, and handwriting recognition. Take information retrieval, for example. If you were to search for "how long is a football bame" (where bame is a medical noun), the search engine would have asked if you had meant "how long is a football game" instead. This is because the probability of "how long is a football bame" is very low according to the language model; in addition, among all of the words easily confused with "bame", "game" would build the most probable sentence.
#### Target Probability
For language model's target probability $P(w_1, ..., w_T)$, if the words in the sentence were to be independent, the joint probability of the whole sentence would be the product of each word's probability:
$$P(w_1, ..., w_T) = \prod_{t=1}^TP(w_t)$$
$$P(w_1, ..., w_T) = \prod_{t=1}^TP(w_t)$$
However we know that the probability of each word depends heavily on previous words, so people usually use the conditional probability to construct the language model:
However, the frequency of words in a sentence typically relates to the words before them, so canonical language models are constructed using conditional probability in its target probability:
In computational linguistics, n-gram is an important text representation method, representing n consecutive items in a text. Based on the desired application scenario, each item could be a letter, a syllable or a word. N-gram model is also an important method in statistical language modeling. When using the n-gram method to train the language model, one uses first (n-1) words to predict the n-th word in a n-gram.
In computational linguistics, n-gram is an important method to represent text. An n-gram represents a contiguous sequence of n consecutive items given a text. Based on the desired application scenario, each item could be a letter, a syllable or a word. The N-gram model is also an important method in statistical language modeling. When training language models with n-grams, the first (n-1) words of an n-gram are used to predict the *n*th word.
Yoshua Bengio and other scientists described how to train a word embedding model using neural network in the famous paper of Neural Probabilistic Language Models \[[1](#reference)\] published in 2003. The Neural Network Language Model (NNLM) described in the paper learns the language model and word embedding simultaneously through a linear transformation and a non-linear hidden connection. By learning from large amount of corpus, we could get the word embedding and then get the probability of the whole sentence through the word embedding. This type of language model can overcome the curse of dimensionality, i.e. model inaccuracy caused by the difference between training and testing data. Caution: because neural network language model is loosely defined, we will not use the name of NNLM but call it N-gram neural model in this chapter.
Yoshua Bengio and other scientists describe how to train a word embedding model using neural network in the famous paper of Neural Probabilistic Language Models \[[1](#reference)\] published in 2003. The Neural Network Language Model (NNLM) described in the paper learns the language model and word embedding simultaneously through a linear transformation and a non-linear hidden connection. That is, after training on large amounts of corpus, the model learns the word embedding; then, it computes the probability of the whole sentence, using the embedding. This type of language model can overcome the **curse of dimensionality** i.e. model inaccuracy caused by the difference in dimensionality between training and testing data. Note that the term *neural network language model* is ill-defined, so we will not use the name NNLM but only refer to it as *N-gram neural model* in this section.
We have previously described to use conditional probability to construct language model, so the probability of the t-th word in a sentence depends on all t-1 words before it. But actually the words further away have less impact on a word, so if we only consider a n-gram, every word is only effected by its previous n-1 words, then we have:
We have previously described language model using conditional probability, where the probability of the *t*-th word in a sentence depends on all $t-1$ words before it. Furthermore, since words further prior have less impact on a word, and every word within an n-gram is only effected by its previous n-1 words, we have:
where $f(w_t, w_{t-1}, ..., w_{t-n+1})$ represents the conditional probability of the current word $w_t$ given its previous n-1 words, and $R(\theta)$ represents parameter regularization term.
where $f(w_t, w_{t-1}, ..., w_{t-n+1})$ represents the conditional probability of the current word $w_t$ given its previous $n-1$ words, and $R(\theta)$ represents parameter regularization term.
<palign="center">
<palign="center">
<imgsrc="image/nnlm.png"width=500><br/>
<imgsrc="image/nnlm.png"width=500><br/>
图2. N-gram neural network model
Figure 2. N-gram neural network model
</p>
</p>
Figure 2 shows the N-gram neural network model. From bottom to top, the model can be separated into the following components:
Figure 2 shows the N-gram neural network model. From the bottom up, the model has the following components:
- For each sample, the model gets input of $w_{t-n+1},...w_{t-1}$, and outputs the probability of the t-th word being one of `|V|` in the dictionary.
- For each sample, the model gets input $w_{t-n+1},...w_{t-1}$, and outputs the probability that the t-th word is one of `|V|` in the dictionary.
Every input word $w_{t-n+1},...w_{t-1}$ first get transformed into word embedding $C(w_{t-n+1}),...C(w_{t-1})$ through matrix multiplication.
Every input word $w_{t-n+1},...w_{t-1}$ first gets transformed into word embedding $C(w_{t-n+1}),...C(w_{t-1})$ through a transformation matrix.
-Then all the word embeddings concatenate into a single vector, which is mapped into the t-th word hidden representation:
-All the word embeddings concatenate into a single vector, which is mapped (nonlinearly) into the $t$-th word hidden representation:
$$g=Utanh(\theta^Tx + b_1) + Wx + b_2$$
$$g=Utanh(\theta^Tx + b_1) + Wx + b_2$$
where $x$ is the large vector concatenated from all the word embeddings representing the context; $\theta$, $U$, $b_1$, $b_2$ and $W$ are parameters connecting from word embedding layer to hidden layer. $g$ represents the unnormalized probability of the output word, $g_i$ represents the unnormalized probability of the output word being the i-th word in the dictionary.
where $x$ is the large vector concatenated from all the word embeddings representing the context; $\theta$, $U$, $b_1$, $b_2$ and $W$ are parameters connecting word embedding layers to the hidden layers. $g$ represents the unnormalized probability of the output word, $g_i$ represents the unnormalized probability of the output word being the i-th word in the dictionary.
- Based on the definition of softmax, by normalizing $g_i$, the probability of the output word being $w_t$ is:
- Based on the definition of softmax, using normalized $g_i$, the probability that the output word is $w_t$ is represented as:
where $y_k^i$ represents the true label for the k-th class in the i-th sample, $softmax(g_k^i)$ represents the softmax probability for the k-th class in the i-th sample.
where $y_k^i$ represents the true label for the $k$-th class in the $i$-th sample ($0$ or $1$), $softmax(g_k^i)$ represents the softmax probability for the $k$-th class in the $i$-th sample.
### Continuous Bag-of-Words model(CBOW)
### Continuous Bag-of-Words model(CBOW)
CBOW model predicts the current word based on N words before and after it. When N=2, the model is as the figure below:
CBOW model predicts the current word based on the N words both before and after it. When $N=2$, the model is as the figure below:
<palign="center">
<palign="center">
<imgsrc="image/cbow.png"width=250><br/>
<imgsrc="image/cbow.png"width=250><br/>
Figure 3. CBOW model
Figure 3. CBOW model
</p>
</p>
Concretely, by ignoring the sequence of words in the context, CBOW uses the average value of the word embedding of the context words to predict the current word:
Specifically, by ignoring the order of words in the sequence, CBOW uses the average value of the word embedding of the context to predict the current word:
where $x_t$ is the word embedding of the t-th word, classification score vector is $z=U*context$, the final classification $y$ uses softmax and the loss function uses multi-class cross-entropy.
where $x_t$ is the word embedding of the t-th word, classification score vector is $z=U*\text{context}$, the final classification $y$ uses softmax and the loss function uses multi-class cross-entropy.
### Skip-gram model
### Skip-gram model
...
@@ -126,11 +143,11 @@ The advantages of CBOW is that it smooths over the word embeddings of the contex
...
@@ -126,11 +143,11 @@ The advantages of CBOW is that it smooths over the word embeddings of the contex
<palign="center">
<palign="center">
<imgsrc="image/skipgram.png"width=250><br/>
<imgsrc="image/skipgram.png"width=250><br/>
图4. Skip-gram model
Figure 4. Skip-gram model
</p>
</p>
As illustrated in the figure above, Skip-gram model maps the word embedding of the given word onto $2n$ word embeddings (including $n$ words before and $n$ words after the given word), and then obtained the classification loss of all those $2n$ words by softmax.
As illustrated in the figure above, skip-gram model maps the word embedding of the given word onto $2n$ word embeddings (including $n$ words before and $n$ words after the given word), and then combine the classification loss of all those $2n$ words by softmax.
## Data Preparation
## Data Preparation
...
@@ -142,7 +159,9 @@ As illustrated in the figure above, Skip-gram model maps the word embedding of t
...
@@ -142,7 +159,9 @@ As illustrated in the figure above, Skip-gram model maps the word embedding of t
## Conclusion
## Conclusion
In this chapter, we introduced word embedding, relationship between language model and word embedding, and how to train neural network model to get word embedding. In the information retrieval, we can obtain the similarity between the query and document keyword by calculating the cosine value between their word embeddings. In the grammar analysis and semantic analysis, a trained word embedding can be used to initialized models to gain better performance. In document classification, we can use clustering method to group synonyms in the documents by word embedding. We hope that readers could use word embedding models in their work after reading this chapter.
This chapter introduces word embedding, the relationship between language model and word embedding, and how to train neural networks to learn word embedding.
In information retrieval, the relevance between the query and document keyword can be computed through the cosine similarity of their word embeddings. In grammar analysis and semantic analysis, a previously trained word embedding can initialize models for better performance. In document classification, clustering the word embedding can group synonyms in the documents. We hope that readers can use word embedding models in their work after reading this chapter.
