/**
* Get all of the words in the evaluation dataset.
*
* @param path
* @return
* @throws Exception
*/
private static Set<String> getWordSimVocab(String path) throws Exception {
Set<String> vocab = new HashSet<String>();
BufferedReader reader = new BufferedReader(new FileReader(path));
String line = "";
line = reader.readLine();
String[] keys = line.split(",");
// Read the first line that contains the column keys.
if (keys.length != 3) {
System.out.println(
"There should be two words per line "
+ "and a single score for each of these word "
+ "pairs. We just saw, "
+ line);
System.exit(0);
}
while ((line = reader.readLine()) != null) {
String[] parts = line.split(",");
if (parts.length != 3) {
System.out.println("WordSim line: " + line + " should contain two words and a score.");
System.exit(0);
}
String word1 = parts[0];
String word2 = parts[1];
vocab.add(word1);
vocab.add(word2);
}
reader.close();
return vocab;
}
/**
* A dumb vector space model that counts each word's co-occurences with a predefined set of
* content words and uses these co-occurence vectors directly as word representations. The context
* in which a word occurs is the set of content words in an entire sentence.
*
* <p>N.B. Most people would probably not consider this an embedding model, since the words have
* not been embedded in a lower dimensional subspace. However, it is a good starting point.
*
* <p>Since this approach does not share any information between representations of different
* words, we can filter the training data to only include sentences that contain words of
* interest. In other approaches this may not be a good idea.
*
* @param dataPath
* @param targetVocab
* @param contentVocab
* @return
*/
private static HashMap<String, float[]> getEmbeddings(
String dataPath, HashMap<String, Integer> contentVocab, Set<String> targetVocab) {
HashMap<String, float[]> embeddingMatrix = new HashMap<String, float[]>();
for (String target_word : targetVocab) {
embeddingMatrix.put(target_word, new float[contentVocab.size()]);
}
Collection<List<String>> sentenceCollection =
SentenceCollection.Reader.readSentenceCollection(dataPath);
for (List<String> sentence : sentenceCollection) {
Set<String> sw = new HashSet<String>(sentence);
sw.retainAll(targetVocab);
for (String word : sentence) {
if (!contentVocab.containsKey(word)) continue;
int contentWordId = contentVocab.get(word);
for (String targetWord : sw) {
embeddingMatrix.get(targetWord)[contentWordId] =
embeddingMatrix.get(targetWord)[contentWordId] + 1;
}
}
}
return embeddingMatrix;
}
private Pair<Integer, Set<Integer>> getWordContextPair(
List<String> sentence, int wordPosition) {
String centerWord = sentence.get(wordPosition);
int centerWordIndex = encodedVocab.get(centerWord);
Set<Integer> contextWordSet = new HashSet<Integer>();
for (int i = wordPosition - contextSize; i < wordPosition + contextSize; i++) {
if (i < 0) continue; // Ignore contexts prior to start of sentence
if (i >= sentence.size()) break; // Ignore contexts after end of current sentence
if (i == centerWordIndex) continue; // Ignore center word
String contextWord = sentence.get(i);
int contextWordIndex = encodedVocab.get(contextWord);
contextWordSet.add(contextWordIndex);
}
return Pair.create(centerWordIndex, contextWordSet);
}
private Set<Integer> negativeSampleContexts(int wordIndex) {
Set<Integer> negativeContexts = new HashSet<Integer>();
Set<Integer> positiveContexts = contextPairs.get(wordIndex);
while (negativeContexts.size() < kSamples) {
int contextIndex = (int) (Math.random() * V);
if (!positiveContexts.contains(contextIndex) && negativeContexts.contains(contextIndex)) {
negativeContexts.add(contextIndex);
}
}
return negativeContexts;
}
/**
* @author jacqueline If boolean sampleUnigram = true, we use noiseSampler from
* randomContextGeneration to model the unigram probability distribution raised to specfied
* power, default 3/4. Otherwise, use overloaded negativeSampleContexts(int wordIndex)
* method to draw from uniform probability distribution.
*/
private Set<Integer> negativeSampleContexts(
int wordIndex, EnumeratedDistribution<String> weightedRandomSample) {
Set<Integer> negativeContexts = new HashSet<Integer>();
Set<Integer> positiveContexts = contextPairs.get(wordIndex);
while (negativeContexts.size() < kSamples) {
String possibleContext = weightedRandomSample.sample();
int contextIndex = encodedVocab.get(possibleContext);
if (!positiveContexts.contains(contextIndex) && !negativeContexts.contains(contextIndex)) {
negativeContexts.add(contextIndex);
}
}
return negativeContexts;
}
public WordSim(
String dataPath,
int kSamples,
int dimensions,
int contextSize,
double power,
double alpha,
double min_eta,
double sigma,
int epochs,
boolean skipGram,
boolean negativeSampling,
boolean sampleUnigram,
double learningDecay) {
this.sentenceCollection = SentenceCollection.Reader.readSentenceCollection(dataPath);
this.kSamples = kSamples;
this.dimensions = dimensions;
this.contextSize = contextSize;
this.power = power;
this.alpha = alpha;
this.min_eta = min_eta;
this.sigma = sigma;
this.epochs = epochs;
this.skipGram = skipGram;
this.negativeSampling = negativeSampling;
this.sampleUnigram = sampleUnigram;
this.learningDecay = learningDecay;
this.vocabulary = LanguageModelTester.extractVocabulary(sentenceCollection);
encodeVocabulary(); // create one-hot encoding index for all words in vocabulary
this.V = vocabulary.size(); // cardinality of vocabulary
setAllContexts(); // create HashMap for all observed positive contexts for each word
if (sampleUnigram)
randomContextGeneration(); // create weighted random sampler for noise distribution
else noiseSampler = null;
}