public static void main(String[] args) throws IOException {
// Parse command line flags and arguments
Map<String, String> argMap = CommandLineUtils.simpleCommandLineParser(args);
// Set up default parameters and settings
String basePath = ".";
String model = "baseline";
boolean verbose = false;
// Update defaults using command line specifications
// The path to the assignment data
if (argMap.containsKey("-path")) {
basePath = argMap.get("-path");
}
System.out.println("Using base path: " + basePath);
// A string descriptor of the model to use
if (argMap.containsKey("-model")) {
model = argMap.get("-model");
}
System.out.println("Using model: " + model);
// Whether or not to print the individual speech errors.
if (argMap.containsKey("-verbose")) {
verbose = true;
}
if (argMap.containsKey("-quiet")) {
verbose = false;
}
// Read in all the assignment data
String trainingSentencesFile = "/treebank-sentences-spoken-train.txt";
String validationSentencesFile = "/treebank-sentences-spoken-validate.txt";
String testSentencesFile = "/treebank-sentences-spoken-test.txt";
String speechNBestListsPath = "/wsj_n_bst";
Collection<List<String>> trainingSentenceCollection =
SentenceCollection.Reader.readSentenceCollection(basePath + trainingSentencesFile);
Collection<List<String>> validationSentenceCollection =
SentenceCollection.Reader.readSentenceCollection(basePath + validationSentencesFile);
Collection<List<String>> testSentenceCollection =
SentenceCollection.Reader.readSentenceCollection(basePath + testSentencesFile);
Set trainingVocabulary = extractVocabulary(trainingSentenceCollection);
List<SpeechNBestList> speechNBestLists =
SpeechNBestList.Reader.readSpeechNBestLists(
basePath + speechNBestListsPath, trainingVocabulary);
// Build the language model
LanguageModel languageModel = null;
if (model.equalsIgnoreCase("baseline")) {
languageModel = new EmpiricalUnigramLanguageModel(trainingSentenceCollection);
} else if (model.equalsIgnoreCase("bigram")) {
languageModel = new EmpiricalBigramLanguageModel(trainingSentenceCollection);
} else if (model.equalsIgnoreCase("trigram")) {
languageModel = new EmpiricalTrigramLanguageModel(trainingSentenceCollection);
} else if (model.equalsIgnoreCase("KN")) {
languageModel = new KNBigramLanguageModel(trainingSentenceCollection);
} else {
throw new RuntimeException("Unknown model descriptor: " + model);
}
boolean tunning = true;
if (tunning == true) {
// Evaluate the language model
File f = new File("3-gram result2.txt");
BufferedWriter output = new BufferedWriter(new FileWriter(f));
double min = 1;
double minR1 = 1;
double minR2 = 1;
double minR3 = 1;
double minR4 = 1;
double minR5 = 1;
double minK = 1;
double minPerp = 1000000;
double k = 0.1;
double step = 0.1;
double r1 = 0;
double r2 = 0;
double r3 = 0;
double r4 = 0;
double r5 = 0;
for (r1 = 0; r1 < 1; r1 += 0.2) {
for (r2 = 0; (r1 + r2) < 1; r2 += 0.1) {
r3 = 1 - (r1 + r2);
for (r4 = 0; r4 < 1; r4 += 0.1) {
r5 = 1 - r4;
EmpiricalTrigramLanguageModel.r1 = r1;
EmpiricalTrigramLanguageModel.r2 = r2;
EmpiricalTrigramLanguageModel.r3 = r3;
EmpiricalTrigramLanguageModel.r4 = r4;
EmpiricalTrigramLanguageModel.r5 = r5;
double wordErrorRate = calculateWordErrorRate(languageModel, speechNBestLists, verbose);
double wsjPerplexity = calculatePerplexity(languageModel, validationSentenceCollection);
double hubPerplexity =
calculatePerplexity(languageModel, extractCorrectSentenceList(speechNBestLists));
if (minPerp > wsjPerplexity) {
minPerp = wsjPerplexity;
// minK = k;
minR1 = r1;
minR2 = r2;
minR3 = r3;
minR4 = r4;
minR5 = r5;
}
if (min > wordErrorRate) min = wordErrorRate;
System.out.println(
"r1:" + r1 + "\t" + "r2:" + r2 + "\t" + "r3:" + r3 + "\t" + "r4:" + r4 + "\t"
+ "r5:" + r5 + "\t");
System.out.println("HUB Word Error Rate: " + wordErrorRate);
System.out.println("Min Error Rate till now: " + min);
// System.out.println("minK=" + minK);
System.out.println("WSJ Perplexity: " + wsjPerplexity);
System.out.println("HUB Perplexity: " + hubPerplexity);
System.out.println();
BigDecimal big_r1 = new BigDecimal(r1);
BigDecimal big_r2 = new BigDecimal(r2);
BigDecimal big_r3 = new BigDecimal(r3);
BigDecimal big_r4 = new BigDecimal(r4);
BigDecimal big_r5 = new BigDecimal(r5);
BigDecimal big_wsjPerplexity = new BigDecimal(wsjPerplexity);
BigDecimal big_hubPerplexity = new BigDecimal(hubPerplexity);
BigDecimal big_wordErrorRate = new BigDecimal(wordErrorRate);
big_r1 = big_r1.setScale(1, BigDecimal.ROUND_HALF_UP);
big_r2 = big_r2.setScale(1, BigDecimal.ROUND_HALF_UP);
big_r3 = big_r3.setScale(1, BigDecimal.ROUND_HALF_UP);
big_r4 = big_r4.setScale(1, BigDecimal.ROUND_HALF_UP);
big_r5 = big_r5.setScale(1, BigDecimal.ROUND_HALF_UP);
big_wsjPerplexity = big_wsjPerplexity.setScale(2, BigDecimal.ROUND_HALF_UP);
big_hubPerplexity = big_hubPerplexity.setScale(2, BigDecimal.ROUND_HALF_UP);
big_wordErrorRate = big_wordErrorRate.setScale(4, BigDecimal.ROUND_HALF_UP);
output.write(
big_r1
+ "\t\t\t"
+ big_r2
+ "\t\t\t"
+ big_r3
+ "\t\t\t"
+ big_r4
+ "\t\t\t"
+ big_r5
+ "\t\t\t"
+ big_wsjPerplexity
+ "\t\t\t"
+ big_hubPerplexity
+ "\t\t\t"
+ big_wordErrorRate);
output.write("\n");
}
}
}
output.write("\n");
output.write("min WER:" + min + "\n");
output.write("min Perp:" + minPerp + "\n");
output.write(
"minR1:" + "\t\t\t" + minR1 + "\t\t\t" + "minR2:" + "\t\t\t" + minR2 + "\t\t\t" + "minR3:"
+ "\t\t\t" + minR3 + "\t\t\t" + "minR4:" + "\t\t\t" + minR4 + "\t\t\t" + "minR5:"
+ "\t\t\t" + minR5 + "\n");
output.close();
} else {
EmpiricalTrigramLanguageModel.k = 0.1;
EmpiricalTrigramLanguageModel.r1 = 0.7;
EmpiricalTrigramLanguageModel.r2 = 0.2;
EmpiricalTrigramLanguageModel.r3 = 0.1;
double wordErrorRate = calculateWordErrorRate(languageModel, speechNBestLists, verbose);
double wsjPerplexity = calculatePerplexity(languageModel, testSentenceCollection);
double hubPerplexity =
calculatePerplexity(languageModel, extractCorrectSentenceList(speechNBestLists));
System.out.println("HUB Word Error Rate: " + wordErrorRate);
System.out.println("WSJ Perplexity: " + wsjPerplexity);
System.out.println("HUB Perplexity: " + hubPerplexity);
}
}
