public static void GenerateFasttextPretrainingFileFromJsonDocuments(string answersDir, string answerUnitsFile, string pretrainingFilePath)
{
var documents = new AnswerStoreSimulator(answerUnitsFile, answersDir);
using (StreamWriter sw = new StreamWriter(pretrainingFilePath, false, Encoding.UTF8))
{
foreach (var paragraph in documents.AnswerUnits)
{
foreach (var title in paragraph.content.title)
{
if (title != "no-title")
{
sw.WriteLine(SentenceClassifier.PreprocessSentence(title));
}
}
sw.WriteLine(SentenceClassifier.PreprocessSentence(paragraph.content.plainText));
}
}
}
public static void GenerateFasttextTrainingFileFromCsvTable(string csvFilePath, string pretrainingFilePath)
{
if (File.Exists(csvFilePath))
{
using (StreamReader sr = new StreamReader(csvFilePath, Encoding.UTF8))
using (StreamWriter sw = new StreamWriter(pretrainingFilePath, false, Encoding.UTF8))
{
string line = null;
while ((line = sr.ReadLine()) != null)
{
int intentIndex = line.LastIndexOf(',');
if (intentIndex < 0)
{
throw new Exception("invalid file format");
}
var question = line.Substring(0, intentIndex);
sw.WriteLine(SentenceClassifier.PreprocessSentence(question));
}
}
}
}
public static void GeneratePerfAnalysisFile()
{
var WORKING_DIR = @"..\fasttext\tests\";
var DATASET_NAME = "SAV_NLC_100520171";
var MODEL_NAME = "model_savings1";
string trainingFilePath = WORKING_DIR + DATASET_NAME + ".train";
ISet <string> intentsSet = new HashSet <string>();
var intentsCountTraining = new Dictionary <string, int>();
var trainingSamples = new Dictionary <string, IList <string> >();
using (StreamReader sr = new StreamReader(trainingFilePath, Encoding.UTF8))
{
string line = null;
while ((line = sr.ReadLine()) != null)
{
int endOfLabel = line.IndexOf(' ');
string intent = line.Substring(9, endOfLabel - 9);
string trainingQuestion = line.Substring(endOfLabel + 1);
if (!intentsSet.Contains(intent))
{
intentsSet.Add(intent);
intentsCountTraining.Add(intent, 1);
}
else
{
intentsCountTraining[intent] += 1;
}
if (!trainingSamples.ContainsKey(intent))
{
trainingSamples.Add(intent, new List <string>());
}
trainingSamples[intent].Add(trainingQuestion);
}
}
string validationFilePath = WORKING_DIR + DATASET_NAME + ".valid";
IDictionary <string, string> annotatedQuestions = new Dictionary <string, string>();
var intentsCountValidation = new Dictionary <string, int>();
using (StreamReader sr = new StreamReader(validationFilePath, Encoding.UTF8))
{
string line = null;
while ((line = sr.ReadLine()) != null)
{
int endOfLabel = line.IndexOf(' ');
string intent = line.Substring(9, endOfLabel - 9);
string question = line.Substring(endOfLabel + 1);
if (!intentsSet.Contains(intent))
{
intentsSet.Add(intent);
}
if (!intentsCountValidation.ContainsKey(intent))
{
intentsCountValidation.Add(intent, 1);
}
else
{
intentsCountValidation[intent] += 1;
}
if (!annotatedQuestions.ContainsKey(question))
{
annotatedQuestions.Add(question, intent);
}
}
}
IList <string> intents = new List <string>(intentsSet.OrderBy(i => i));
IDictionary <string, int> intentsIndexes = new Dictionary <string, int>();
for (int i = 0; i < intents.Count; i++)
{
var intent = intents[i];
intentsIndexes.Add(intent, i);
}
int[,] confusionMatrix = new int[intents.Count, intents.Count];
var predictionResults = new List <PredictionResult>();
using (SentenceClassifier classifier = new SentenceClassifier(WORKING_DIR + MODEL_NAME))
{
foreach (var question in annotatedQuestions.Keys)
{
var annotatedIntent = annotatedQuestions[question];
var annotatedIntentIndex = intentsIndexes[annotatedIntent];
var result = classifier.PredictLabels(question);
var predictionResult = new PredictionResult();
predictionResult.Question = question;
predictionResult.ExpectedIntentIndex = annotatedIntentIndex;
predictionResult.IntentIndex1 = intentsIndexes[result.Label1];
predictionResult.IntentProba1 = result.Proba1;
if (result.Label2 != null)
{
predictionResult.IntentIndex2 = intentsIndexes[result.Label2];
predictionResult.IntentProba2 = result.Proba2;
}
predictionResults.Add(predictionResult);
confusionMatrix[annotatedIntentIndex, predictionResult.IntentIndex1]++;
}
}
var intentsPerfs = new List <IntentPerf>();
for (int intentIndex = 0; intentIndex < intents.Count; intentIndex++)
{
string intent = intents[intentIndex];
int truePositives = confusionMatrix[intentIndex, intentIndex];
int falsePositives = 0;
for (int j = 0; j < intents.Count; j++)
{
if (j != intentIndex)
{
falsePositives += confusionMatrix[j, intentIndex];
}
}
int falseNegatives = 0;
for (int j = 0; j < intents.Count; j++)
{
if (j != intentIndex)
{
falseNegatives += confusionMatrix[intentIndex, j];
}
}
var intentPerf = new IntentPerf();
intentPerf.Intent = intent;
if ((truePositives + falsePositives) > 0)
{
intentPerf.Precision = (float)truePositives / (truePositives + falsePositives);
}
else
{
intentPerf.Precision = -1;
}
if ((truePositives + falseNegatives) > 0)
{
intentPerf.Recall = (float)truePositives / (truePositives + falseNegatives);
}
else
{
intentPerf.Recall = -1;
}
if (intentPerf.Precision > 0 && intentPerf.Recall > 0)
{
intentPerf.F1 = 2 * intentPerf.Precision * intentPerf.Recall / (intentPerf.Precision + intentPerf.Recall);
}
else
{
intentPerf.F1 = -1;
}
intentsPerfs.Add(intentPerf);
}
var intentsConfusion = new List <IntentsPairConfusion>();
for (int expectedIntentIndex = 0; expectedIntentIndex < intents.Count; expectedIntentIndex++)
{
for (int predictedIntentIndex = 0; predictedIntentIndex < expectedIntentIndex; predictedIntentIndex++)
{
var intentsPair = new IntentsPairConfusion();
intentsPair.IntentIndex1 = expectedIntentIndex;
intentsPair.IntentIndex2 = predictedIntentIndex;
intentsPair.Expected1Found1 = confusionMatrix[expectedIntentIndex, expectedIntentIndex];
intentsPair.Expected1Found2 = confusionMatrix[expectedIntentIndex, predictedIntentIndex];
intentsPair.Expected2Found1 = confusionMatrix[predictedIntentIndex, expectedIntentIndex];
intentsPair.Expected2Found2 = confusionMatrix[predictedIntentIndex, predictedIntentIndex];
intentsPair.ConfusionErrors = intentsPair.Expected1Found2 + intentsPair.Expected2Found1;
if (intentsPair.Expected1Found1 > 0)
{
intentsPair.ConfusionRate1To2 = (float)intentsPair.Expected1Found2 / intentsPair.Expected1Found1;
}
else if (intentsPair.Expected1Found2 > 0)
{
intentsPair.ConfusionRate1To2 = 1;
}
if (intentsPair.Expected2Found2 > 0)
{
intentsPair.ConfusionRate2To1 = (float)intentsPair.Expected2Found1 / intentsPair.Expected2Found2;
}
else if (intentsPair.Expected2Found1 > 0)
{
intentsPair.ConfusionRate2To1 = 1;
}
intentsConfusion.Add(intentsPair);
}
}
using (StreamWriter sw = new StreamWriter(@"..\fasttext\" + MODEL_NAME + ".results.csv", false, Encoding.GetEncoding("iso8859-1")))
{
sw.WriteLine("1. Intents performance");
sw.WriteLine();
sw.WriteLine("Intent;# Training;# Validation;F1;Precision;Recall");
foreach (var intentPerf in intentsPerfs.OrderByDescending(perf => perf.F1))
{
var trainingCount = 0;
intentsCountTraining.TryGetValue(intentPerf.Intent, out trainingCount);
var validationCount = 0;
intentsCountValidation.TryGetValue(intentPerf.Intent, out validationCount);
sw.Write(intentPerf.Intent);
sw.Write(';');
sw.Write(trainingCount);
sw.Write(';');
sw.Write(validationCount);
sw.Write(';');
sw.Write(intentPerf.F1 < 0 ? "N/A" : intentPerf.F1.ToString("N2"));
sw.Write(';');
sw.Write(intentPerf.Precision < 0 ? "N/A" : intentPerf.Precision.ToString("N2"));
sw.Write(';');
sw.Write(intentPerf.Recall < 0 ? "N/A" : intentPerf.Recall.ToString("N2"));
sw.WriteLine();
}
sw.WriteLine();
sw.WriteLine("2. Intents confusion (top 50)");
sw.WriteLine();
sw.WriteLine("Intent 1;Intent 2;# Confusions;# Confusion 1>2;% Confusion 1>2;# Confusion 2>1;% Confusion 2>1");
foreach (var intentPair in intentsConfusion.OrderByDescending(pair => pair.ConfusionErrors).Take(50))
{
sw.Write(intents[intentPair.IntentIndex1]);
sw.Write(';');
sw.Write(intents[intentPair.IntentIndex2]);
sw.Write(';');
sw.Write(intentPair.ConfusionErrors);
sw.Write(';');
sw.Write(intentPair.Expected1Found2);
sw.Write(';');
sw.Write(intentPair.ConfusionRate1To2.ToString("N2"));
sw.Write(';');
sw.Write(intentPair.Expected2Found1);
sw.Write(';');
sw.Write(intentPair.ConfusionRate2To1.ToString("N2"));
sw.WriteLine();
}
sw.WriteLine();
sw.WriteLine();
sw.WriteLine();
sw.WriteLine("3. Detailed error analysis for each intent");
sw.WriteLine();
for (int expectedIntentIndex = 0; expectedIntentIndex < intents.Count; expectedIntentIndex++)
{
var expectedIntent = intents[expectedIntentIndex];
sw.WriteLine(">>> " + expectedIntent);
sw.WriteLine();
sw.WriteLine("Training set questions");
if (trainingSamples.ContainsKey(expectedIntent))
{
foreach (var trainingQuestion in trainingSamples[expectedIntent])
{
sw.WriteLine(trainingQuestion);
}
}
else
{
sw.WriteLine("-- NONE --");
}
sw.WriteLine();
sw.WriteLine("Questions correctly classified in this class");
foreach (var predictionResult in predictionResults.Where(pred => pred.ExpectedIntentIndex == expectedIntentIndex && pred.IntentIndex1 == expectedIntentIndex).OrderByDescending(pred => pred.IntentProba1))
{
sw.Write(predictionResult.Question.Replace(';', ' '));
sw.Write(';');
sw.Write(predictionResult.IntentProba1);
sw.Write(';');
sw.Write(predictionResult.IntentProba1 - predictionResult.IntentProba2);
sw.WriteLine();
}
sw.WriteLine();
sw.WriteLine("Questions from this class incorrectly classified in another class");
foreach (var predictionResult in predictionResults.Where(pred => pred.ExpectedIntentIndex == expectedIntentIndex && pred.IntentIndex1 != expectedIntentIndex).OrderBy(pred => pred.IntentIndex1).ThenByDescending(pred => pred.IntentProba1))
{
sw.Write(predictionResult.Question.Replace(';', ' '));
sw.Write(';');
sw.Write(intents[predictionResult.IntentIndex1]);
sw.Write(';');
sw.Write(predictionResult.IntentProba1);
sw.Write(';');
sw.Write(predictionResult.IntentProba1 - predictionResult.IntentProba2);
sw.Write(';');
sw.Write(intents[predictionResult.IntentIndex2]);
sw.Write(';');
sw.Write(predictionResult.IntentProba2);
sw.WriteLine();
}
sw.WriteLine();
sw.WriteLine("Questions from another class incorrectly classified in this class");
foreach (var predictionResult in predictionResults.Where(pred => pred.ExpectedIntentIndex != expectedIntentIndex && pred.IntentIndex1 == expectedIntentIndex).OrderBy(pred => pred.ExpectedIntentIndex).ThenByDescending(pred => pred.IntentProba1))
{
sw.Write(predictionResult.Question.Replace(';', ' '));
sw.Write(';');
sw.Write(intents[predictionResult.ExpectedIntentIndex]);
sw.Write(';');
sw.Write(predictionResult.IntentProba1);
sw.Write(';');
sw.Write(predictionResult.IntentProba1 - predictionResult.IntentProba2);
sw.Write(';');
sw.Write(intents[predictionResult.IntentIndex2]);
sw.Write(';');
sw.Write(predictionResult.IntentProba2);
sw.WriteLine();
}
sw.WriteLine();
}
}
}
public static void GenerateFasttextTrainingFileFromCsvTable(string csvFilePath, int splitTrainingSets)
{
if (File.Exists(csvFilePath))
{
Console.WriteLine("Reading file : " + csvFilePath + " ...");
int lineCount = 0;
var questions = new List <LabelAndQuestion>();
using (StreamReader sr = new StreamReader(csvFilePath, Encoding.UTF8))
{
string line = null;
while ((line = sr.ReadLine()) != null)
{
int intentIndex = line.LastIndexOf(',');
if (intentIndex < 0)
{
throw new Exception("invalid file format");
}
var labelAndQuestion = new LabelAndQuestion();
labelAndQuestion.Question = line.Substring(0, intentIndex);
labelAndQuestion.Label = line.Substring(intentIndex + 1);
questions.Add(labelAndQuestion);
}
}
Shuffle(questions);
int bucketQuestionsCount = questions.Count / splitTrainingSets;
string csvFileName = Path.GetFileNameWithoutExtension(csvFilePath);
string csvFileDirectory = Path.GetDirectoryName(csvFilePath);
for (int trainingSetNumber = 1; trainingSetNumber <= splitTrainingSets; trainingSetNumber++)
{
string trainingFilePath = csvFileDirectory + Path.DirectorySeparatorChar + csvFileName + trainingSetNumber + ".train";
string validationFilePath = csvFileDirectory + Path.DirectorySeparatorChar + csvFileName + trainingSetNumber + ".valid";
if (splitTrainingSets == 1)
{
trainingFilePath = csvFileDirectory + Path.DirectorySeparatorChar + csvFileName + ".delete";
validationFilePath = csvFileDirectory + Path.DirectorySeparatorChar + csvFileName + ".train";
}
using (StreamWriter trainsw = new StreamWriter(trainingFilePath, false, Encoding.UTF8))
{
using (StreamWriter validsw = new StreamWriter(validationFilePath, false, Encoding.UTF8))
{
for (int questionIndex = 0; questionIndex < questions.Count; questionIndex++)
{
var labelAndQuestion = questions[questionIndex];
StringBuilder sbQuestion = new StringBuilder();
sbQuestion.Append(FASTTEXT_LABEL_PREFIX);
sbQuestion.Append(labelAndQuestion.Label);
sbQuestion.Append(' ');
sbQuestion.Append(SentenceClassifier.PreprocessSentence(labelAndQuestion.Question));
bool writeToValidation = questionIndex >= (trainingSetNumber - 1) * bucketQuestionsCount && questionIndex < trainingSetNumber * bucketQuestionsCount;
if (!writeToValidation)
{
trainsw.WriteLine(sbQuestion.ToString());
}
else
{
validsw.WriteLine(sbQuestion.ToString());
}
lineCount++;
}
}
}
Console.WriteLine("OK - " + lineCount + " training samples written to " + trainingFilePath);
}
}
else
{
Console.WriteLine("ERROR : File " + csvFilePath + " doesn't exist");
}
}
