public static void prepareWord2VecCorpus(
Properties properties, Parameters parameters, String output) throws Exception {
File fAbstractDir = new File(PropertiesUtils.getString(properties, "corpusDir", ""));
File fOutputFile = new File(output);
Tokenizer tokenizer = new Tokenizer(true, ' ');
MaxentTagger tagger = new MaxentTagger(PropertiesUtils.getString(properties, "pos_tagger", ""));
Morphology morphology = new Morphology();
OutputStreamWriter osw = new OutputStreamWriter(new FileOutputStream(fOutputFile), "utf-8");
for (File abstractFile : fAbstractDir.listFiles()) {
Abstract ab = (Abstract) ObjectSerializer.readObjectFromFile(abstractFile.getAbsolutePath());
for (ADESentence sentence : ab.sentences) {
List<CoreLabel> tokens = tokenizer.tokenize(sentence.offset, sentence.text);
tagger.tagCoreLabels(tokens);
for (int i = 0; i < tokens.size(); i++) morphology.stem(tokens.get(i));
for (CoreLabel token : tokens) {
String temp = NNADE.wordPreprocess(token, parameters);
osw.write(temp + " ");
}
osw.write("\n");
}
}
osw.close();
System.out.println("Generate a word2vec corpus.");
System.out.printf("wordPreprocess = %d%n", parameters.wordPreprocess);
System.out.printf("embeddingSize = %d%n", parameters.embeddingSize);
}
static void findClosestWord(String w) throws Exception {
NNADE nnade = (NNADE) ObjectSerializer.readObjectFromFile("E:\\ade\\nnade.ser0");
ArrayList<String> words = new ArrayList<>();
ArrayList<Matrix> vectors = new ArrayList<>();
Matrix wVector = null;
for (String word : nnade.knownWords) {
int index = nnade.wordIDs.get(word);
if (index == 0 || index == 1) {
System.out.println(word);
continue;
}
if (w.equals(word)) {
wVector = new Matrix(1, 50, nnade.E[index]);
continue;
}
words.add(word);
vectors.add(new Matrix(1, 50, nnade.E[index]));
}
// normalize
/*for(int i=0;i<vectors.size();i++)
Normalizer.doVectorNormalizing(vectors.get(i));*/
int max = 100;
ArrayList<Matrix> mat = new ArrayList<>();
ArrayList<String> word = new ArrayList<>();
for (int i = 0; i < vectors.size(); i++) {
double currentDist = Matrix.distanceEuclidean(vectors.get(i), wVector);
boolean inserted = false;
for (int j = 0; j < mat.size(); j++) {
double oldDist = Matrix.distanceEuclidean(mat.get(j), wVector);
if (currentDist < oldDist) {
mat.add(j, vectors.get(i));
word.add(j, words.get(i));
inserted = true;
break;
}
}
if (inserted == false) {
mat.add(vectors.get(i));
word.add(words.get(i));
}
if (mat.size() > max) {
mat.remove(mat.size() - 1);
word.remove(word.size() - 1);
}
}
// dump result
for (String t : word) {
System.out.print(t + " ");
}
}
static void kmeansWord(Tool tool, Parameters parameters, List<Abstract> abs) throws Exception {
NNADE nnade = (NNADE) ObjectSerializer.readObjectFromFile("E:\\ade\\nnade.ser0");
HashSet<String> results =
new HashSet<String>(
Arrays.asList(
new String[] {
"induce",
"clozapine",
"mtx",
"heparin",
"colchicine",
"methotrexate",
"thrombocytopenia",
"glaucoma",
"myopathy",
"thyrotoxicosis"
}));
/*HashSet<String> results = new HashSet<String>();
HashSet<String> triggers = new HashSet<String>(Arrays.asList(new String[]
{"induce", "associate", "relate", "cause", "develop", "produce", "after", "follow", "result"}));
nnade.nn.preCompute();
for(Abstract ab:abs) {
for(ADESentence gold:ab.sentences) {
List<CoreLabel> tokens = nnade.prepareNLPInfo(tool, gold);
ADESentence predicted = null;
predicted = nnade.decode(tokens, tool);
for(RelationEntity relation:predicted.relaitons) {
Entity former = relation.getFormer();
if(former.start!=former.end)
continue;
Entity latter = relation.getLatter();
if(latter.start!=latter.end)
continue;
for(int i=former.end+1;i<=latter.start-1;i++) {
if(triggers.contains(tokens.get(i).lemma().toLowerCase())) {
results.add(tokens.get(former.start).lemma().toLowerCase());
results.add(tokens.get(latter.start).lemma().toLowerCase());
results.add(tokens.get(i).lemma().toLowerCase());
}
}
}
}
}*/
ArrayList<String> words = new ArrayList<>();
ArrayList<Matrix> vectors = new ArrayList<>();
for (String word : results) {
int index = nnade.wordIDs.get(word);
if (index == 0 || index == 1) {
System.out.println(word);
continue;
}
words.add(word);
vectors.add(new Matrix(1, 50, nnade.E[index]));
}
int k = 2;
// normalize
for (int i = 0; i < vectors.size(); i++) Normalizer.doVectorNormalizing(vectors.get(i));
// do Buckshot
Buckshot bs = new Buckshot(k, vectors);
ArrayList<Matrix> centroids = bs.doBuckshot();
// do KMeans
KMeans mk = new KMeans(k, vectors, centroids, 1000);
mk.getResults();
OutputStreamWriter osw1 =
new OutputStreamWriter(new FileOutputStream("E:\\ade\\kmeans.txt"), "utf-8");
// dump result, i denotes class, j denotes word
for (int i = 0; i < k; i++) {
osw1.write("##the " + i + " class :\n");
int line = 0;
for (int j = 0; j < mk.vectors2classes.length; j++) {
if (i == mk.vectors2classes[j]) {
osw1.write(words.get(j) + " ");
line++;
if (line == 10) {
osw1.write("\n");
line = 0;
}
}
}
osw1.write("\n\n");
}
osw1.close();
int max = 10;
ArrayList<ArrayList<Matrix>> outMat = new ArrayList<ArrayList<Matrix>>();
ArrayList<ArrayList<String>> outWord = new ArrayList<ArrayList<String>>();
for (int i = 0; i < k; i++) {
outMat.add(new ArrayList<Matrix>());
outWord.add(new ArrayList<String>());
}
for (int i = 0; i < mk.vectors2classes.length; i++) {
ArrayList<Matrix> mat = outMat.get(mk.vectors2classes[i]);
ArrayList<String> word = outWord.get(mk.vectors2classes[i]);
double currentDist = mk.similarity(vectors.get(i), mk.centroids.get(mk.vectors2classes[i]));
boolean inserted = false;
for (int j = 0; j < mat.size(); j++) {
double oldDist = mk.similarity(mat.get(j), mk.centroids.get(mk.vectors2classes[i]));
if (currentDist < oldDist) {
mat.add(j, vectors.get(i));
word.add(j, words.get(i));
inserted = true;
break;
}
}
if (inserted == false) {
mat.add(vectors.get(i));
word.add(words.get(i));
}
if (mat.size() > max) {
mat.remove(mat.size() - 1);
word.remove(word.size() - 1);
}
}
// dump result, i denotes class, j denotes word
for (int i = 0; i < outWord.size(); i++) {
ArrayList<String> word = outWord.get(i);
System.out.print("##the " + (i + 1) + " class :\n");
for (String w : word) {
System.out.print(w + " ");
}
System.out.println();
}
}
