smile.nlp.embedding.Word2Vec Maven / Gradle / Ivy
/*
* Copyright (c) 2010-2021 Haifeng Li. All rights reserved.
*
* Smile is free software: you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* Smile is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with Smile. If not, see .
*/
package smile.nlp.embedding;
import java.io.FileInputStream;
import java.io.IOException;
import java.nio.ByteOrder;
import java.nio.FloatBuffer;
import java.nio.MappedByteBuffer;
import java.nio.file.Path;
import java.nio.channels.FileChannel;
import java.util.HashMap;
import java.util.stream.IntStream;
import smile.data.DataFrame;
import smile.data.vector.FloatVector;
/**
* Word2vec is a group of related models that are used to produce word
* embeddings. These models are shallow, two-layer neural networks that
* are trained to reconstruct linguistic contexts of words. Word2vec
* takes as its input a large corpus of text and produces a vector space,
* typically of several hundred dimensions, with each unique word in the
* corpus being assigned a corresponding vector in the space. Word vectors
* are positioned in the vector space such that words that share common
* contexts in the corpus are located close to one another in the space.
*
* Word2vec can utilize either of two model architectures to produce
* a distributed representation of words: continuous bag-of-words (CBOW)
* or continuous skip-gram. In the continuous bag-of-words architecture,
* the model predicts the current word from a window of surrounding context
* words. The order of context words does not influence prediction
* (bag-of-words assumption). In the continuous skip-gram architecture,
* the model uses the current word to predict the surrounding window of
* context words. The skip-gram architecture weighs nearby context words
* more heavily than more distant context words. According to the authors'
* note, CBOW is faster while skip-gram is slower but does a better job
* for infrequent words.
*
* @author Haifeng Li
*/
public class Word2Vec {
/** The vocabulary. */
public final String[] words;
/** The vector space. */
public final DataFrame vectors;
/** The word-to-index map. */
private final HashMap map;
/**
* Constructor.
* @param words the vocabulary.
* @param vectors the vectors of d x n, where d is the dimension
* and n is the size of vocabulary.
*/
public Word2Vec(String[] words, float[][] vectors) {
this.words = words;
this.vectors = DataFrame.of(
IntStream.range(0, vectors.length)
.mapToObj(i -> FloatVector.of("V"+(i+1), vectors[i]))
.toArray(FloatVector[]::new)
);
int n = words.length;
map = new HashMap<>(n * 4 / 3 + 3);
for (int i = 0; i < n; i++) {
map.put(words[i], i);
}
}
/**
* Returns the dimension of embedding vector space.
* @return the dimension of embedding vector space.
*/
public int dimension() {
return vectors.ncol();
}
/**
* Returns the embedding vector of a word.
* @param word the word.
* @return the embedding vector.
*/
public float[] get(String word) {
Integer index = map.get(word);
if (index == null) return null;
int i = index;
int dim = vectors.ncol();
float[] vector = new float[dim];
for (int j = 0; j < dim; j++) {
vector[j] = vectors.getFloat(i, j);
}
return vector;
}
/**
* Returns the embedding vector of a word. For Scala convenience.
* @param word the word.
* @return the embedding vector.
*/
public float[] apply(String word) {
return get(word);
}
/**
* Loads a pre-trained
* word2vec model from binary file of ByteOrder.LITTLE_ENDIAN.
* @param file the path to model file.
* @throws IOException when fails to read the file.
* @return the word2vec model.
*/
public static Word2Vec of(Path file) throws IOException {
return of(file, ByteOrder.LITTLE_ENDIAN);
}
/**
* Loads a pre-trained
* word2vec model from binary file.
* @param file the path to model file.
* @param order the byte order of model file.
* @throws IOException when fails to read the file.
* @return the word2vec model.
*/
public static Word2Vec of(Path file, ByteOrder order) throws IOException {
final long GB = 1024 * 1024 * 1024;
try (FileInputStream input = new FileInputStream(file.toFile())) {
FileChannel channel = input.getChannel();
MappedByteBuffer buffer = channel.map(FileChannel.MapMode.READ_ONLY, 0,
Math.min(channel.size(), Integer.MAX_VALUE));
buffer.order(order);
// Java memory-mapping is up to 2GB. Map chunk per 1GB.
int blocks = 1;
StringBuilder sb = new StringBuilder();
char c = (char) buffer.get();
while (c != '\n') {
sb.append(c);
c = (char) buffer.get();
}
String line = sb.toString();
String[] tokens = line.split("\\s+");
if (tokens.length != 2) {
throw new IllegalStateException("Invalid first line: " + line);
}
int size = Integer.parseInt(tokens[0]);
int dim = Integer.parseInt(tokens[1]);
String[] words = new String[size];
float[][] vectors = new float[dim][size];
for (int i = 0; i < size; i++) {
// read vocab
sb.setLength(0);
c = (char) buffer.get();
while (c != ' ') {
// some binary files have newline
if (c != '\n') sb.append(c);
c = (char) buffer.get();
}
// some binary files have newline
words[i] = sb.toString();
// read vector
FloatBuffer floatBuffer = buffer.asFloatBuffer();
for (int j = 0; j < dim; j++) {
vectors[j][i] = floatBuffer.get();
}
buffer.position(buffer.position() + 4 * dim);
// remap file
if (buffer.position() > GB) {
int newPosition = (int) (buffer.position() - GB);
long chunk = Math.min(channel.size() - GB * blocks, Integer.MAX_VALUE);
buffer = channel.map(FileChannel.MapMode.READ_ONLY, GB * blocks, chunk);
buffer.order(order);
buffer.position(newPosition);
blocks += 1;
}
}
return new Word2Vec(words, vectors);
}
}
}