public TextSource Process(IEnumerable<string> documents, string name)
{
TextSource result = new TextSource();
foreach (var item in documents)
{
var doc = new DocumentSource();
var sItem = item.Replace(".", " . ")
.Replace(",", " , ")
.Replace("\r\n", " \r##n ")
.Replace("\n", " \n ")
.Replace(",", " , ")
.Replace("##n", "\n");
// consider punctuation marks/new line as seperate words
foreach (var segment in sItem.Split(new[] { ' ' }, StringSplitOptions.RemoveEmptyEntries))
{
doc.LanguageSegments.Add(segment);
}
result.Documents.Add(doc);
}
result.Name = name;
return result;
}
public void Init(TextSource referenceSource, int n)
{
if (_initPassed) return;
_initPassed = true;
if (n < 1)
{
_b = 0;
return;
}
_referenceSource = referenceSource;
int n1 = 0;
int n2 = 0;
IEnumerable<NGramCache> nGrams = referenceSource.GetNGramCache().NextSegment.Values;
while (n > 1)
{
n--;
nGrams = nGrams.SelectMany(el => el.NextSegment.Values);
}
foreach (var item in nGrams)
{
var frequency = Convert.ToInt32(item.Value);
if (frequency == 1) n1++;
if (frequency == 2) n2++;
}
_b = (double)n1 / ((double)n1 + 2 * (double)n2);
}
public void Init(TextSource referenceSource, int n)
{
if (_initPassed) return;
_initPassed = true;
if (n < 1)
{
_factor = 1;
return;
}
int n1 = 0;
IEnumerable<NGramCache> nGrams = referenceSource.GetNGramCache().NextSegment.Values;
while (n > 1)
{
n--;
nGrams = nGrams.SelectMany(el => el.NextSegment.Values);
}
foreach (var item in nGrams)
{
var frequency = Convert.ToInt32(item.Value);
if (frequency == 1) n1++;
}
_factor = 1.0 - ((double)n1 / (double)referenceSource.GetAllSegments().Count());
}
public TextSource Process(IEnumerable<string> documents, string name)
{
TextSource result = new TextSource();
foreach (var item in documents)
{
var doc = new DocumentSource();
foreach (var segment in item)
{
doc.LanguageSegments.Add(segment.ToString());
}
result.Documents.Add(doc);
}
result.Name = name;
return result;
}
public void Start(string mainFolder, string outputDir, ISmoothingTechnique smoothing, int nGramSize, ILanguageProcessor processor)
{
var resultSet = new ResultSet();
var bayesClassifier = new BayesTextClassifier();
var docReader = new ReadDocumentFromTiraFile();
var categories = new List<TextSource>();
Console.WriteLine("Scanning...");
dynamic jsonConfig;
using (StreamReader sr = new StreamReader(Path.Combine(mainFolder, "meta-file.json")))
{
jsonConfig = JsonConvert.DeserializeObject(sr.ReadToEnd());
}
var unknownFolder = (string)jsonConfig.folder;
var encodingString = (string)jsonConfig.encoding;
var cultureString = (string)jsonConfig.language;
CultureInfo ci = null;
switch (cultureString)
{
case "EN":
ci = new CultureInfo("en-US");
break;
default:
throw new ApplicationException("culture not found");
}
Encoding encoding = null;
switch (encodingString)
{
case "UTF8":
encoding = Encoding.UTF8;
break;
case "ASCII":
encoding = Encoding.ASCII;
break;
default:
throw new ApplicationException("encoding not found");
}
foreach(var item in jsonConfig["candidate-authors"])
{
var authorName = (string)item["author-name"];
var docs = new DirectoryInfo(Path.Combine(mainFolder, authorName)).GetFiles();
var dataSource = new List<string>();
foreach (var doc in docs)
{
try
{
dataSource.Add(docReader.ReadDocumentText(doc.FullName, encoding, ci));
}
catch
{
Console.WriteLine("Document {0} unreadable", doc.FullName);
}
}
categories.Add(processor.Process(dataSource, authorName));
}
int n = nGramSize;
Console.WriteLine("Scanned {1} documents in {0} categories", categories.Count, categories.Select(el => el.Documents.Count).Aggregate((el1, el2) => el1 + el2));
var allInOne = new TextSource();
allInOne.Documents.AddRange(categories.SelectMany(el => el.Documents));
Console.WriteLine("Building hash tables ..", n);
Parallel.ForEach(categories, category =>
{
for (int i = 1; i <= n; i++)
{
category.BuildSegmentTable(i);
Console.WriteLine("hashed {0} with n={1}", category.Name, i);
}
});
allInOne.SetNGramCache(NGramCache.Aggregate(categories.Select(el => el.GetNGramCache())));
Console.WriteLine("aggregated hashing");
Console.WriteLine("Getting smoothing ready ..");
smoothing.Init(allInOne, nGramSize);
var categoriesToTest = new Dictionary<TextSource, CategoryProbabilityDistribution>();
foreach (var cat in categories)
{
categoriesToTest[cat] = new CategoryProbabilityDistribution(cat, smoothing, n);
}
Console.WriteLine("Start classifying ..");
int totalProgress = jsonConfig["unknown-texts"].Count * categoriesToTest.Count;
int progress = 0;
System.Timers.Timer t = new System.Timers.Timer(5000);
t.Elapsed += (sender, eventArgs) =>
{
Console.Title = "Task is Running. Progress: " + Math.Round((((double)progress / (double)totalProgress) * 100.0), 2).ToString() + "%";
};
t.AutoReset = true;
t.Start();
foreach (var item in jsonConfig["unknown-texts"])
{
TextSource topCategory = null;
var maxProb = 0.0;
var textName = (string)item["unknown-text"];
var probs = new List<double>();
Parallel.ForEach(categoriesToTest, catDist =>
{
var docText = new string[] { docReader.ReadDocumentText(Path.Combine(mainFolder, unknownFolder, textName), encoding, ci) };
var docSource = processor.Process(docText, "unknown").Documents.First();
double p = bayesClassifier.P_c(catDist.Value, docSource, n, (double)catDist.Key.Documents.Count / (double)allInOne.Documents.Count);
lock (probs)
{
probs.Add(p);
if (topCategory == null || p > maxProb)
{
topCategory = catDist.Key;
maxProb = p;
}
}
Interlocked.Increment(ref progress);
});
// getting the score
probs.Remove(maxProb);
double pre_score = 0.0;
double max_sub_score = 0.0;
foreach (var p in probs)
{
var subScore = Math.Abs((maxProb - p) / maxProb) * Math.Pow(Math.E, 3); // normalized difference
var eSubScore = Math.Exp(-subScore);
pre_score += eSubScore;
if (eSubScore > max_sub_score)
max_sub_score = eSubScore;
}
double score = Math.Round(1.0 - (0.5 * (pre_score / probs.Count) + 0.5 * max_sub_score), 2);
Console.WriteLine("Classified {0} as author {1} with score {2}", textName, topCategory.Name, score);
resultSet.answers.Add(new Result(textName, topCategory.Name, score));
Console.WriteLine("writing data to file ...");
string data = JsonConvert.SerializeObject(resultSet, Formatting.Indented);
using (StreamWriter sw = new StreamWriter(Path.Combine(outputDir, "answers.json"), false))
{
sw.Write(data);
sw.Flush();
}
}
}
public void Start(ILanguageProcessor processor, ISmoothingTechnique smoothing)
{
var baseDirectory = AppDomain.CurrentDomain.BaseDirectory;
var bayesClassifier = new BayesTextClassifier();
var docReader = new ReadDocumentFromXmlFile();
var docPath = Path.Combine(baseDirectory, "data");
var authors = new DirectoryInfo(docPath).GetDirectories();
var categories = new List<TextSource>();
// Prepare data
foreach (var item in authors)
{
var docs = item.GetFiles();
var dataSource = new List<string>();
foreach (var doc in docs)
{
try
{
dataSource.Add(docReader.ReadDocumentText(doc.FullName, Encoding.GetEncoding(1253), new CultureInfo("el-GR")));
}
catch
{
Console.WriteLine("Document {0} unreadable", doc.FullName);
}
}
categories.Add(processor.Process(dataSource, item.Name));
}
Console.WriteLine("Scanned {1} documents in {0} categories", categories.Count, categories.Select(el => el.Documents.Count).Aggregate((el1, el2) => el1 + el2));
var testPath = Path.Combine(baseDirectory, "test");
var testAuthors = new DirectoryInfo(testPath).GetDirectories();
var allInOne = new TextSource();
allInOne.Documents.AddRange(categories.SelectMany(el => el.Documents));
// choose n from 1 to 4
for (int n = 1; n <= 4; n++)
{
Console.WriteLine("-----PREPARE for n = {0}", n);
Console.WriteLine("Building hash tables ..", n);
Parallel.ForEach(categories, category =>
{
category.BuildSegmentTable(n);
});
allInOne.SetNGramCache(NGramCache.Aggregate(categories.Select(el => el.GetNGramCache())));
Console.WriteLine("Getting smoothing ready ..");
smoothing.Init(allInOne, n);
var categoriesToTest = new Dictionary<TextSource, CategoryProbabilityDistribution>();
foreach(var cat in categories)
{
categoriesToTest[cat] = new CategoryProbabilityDistribution(cat, smoothing, n);
}
int rightClassified = 0;
int wrongClassified = 0;
Console.WriteLine("-----Algorithm starts now");
foreach (var testAuthor in testAuthors)
{
foreach (var testDocument in testAuthor.GetFiles())
{
TextSource topCategory = null;
var maxProb = 0.0;
Parallel.ForEach(categoriesToTest, catDist =>
{
var docText = new[] { docReader.ReadDocumentText(testDocument.FullName, Encoding.GetEncoding(1253), new CultureInfo("el-GR")) };
var docSource = processor.Process(docText, testAuthor.Name).Documents.First();
double p = bayesClassifier.P_c(catDist.Value, docSource, n, 1.0 / (double)categories.Count);
if (topCategory == null || p > maxProb)
{
topCategory = catDist.Key;
maxProb = p;
}
});
Console.WriteLine("Classified {0} as author {1} - {2}", testDocument.Name, topCategory.Name, topCategory.Name == testAuthor.Name ? "correct" : "incorrect");
if (topCategory.Name == testAuthor.Name) rightClassified++;
else wrongClassified++;
}
}
Console.WriteLine("-----SUMMARY");
Console.WriteLine("Success rate for n={0} is {1}\n", n, (double)rightClassified / ((double)rightClassified + (double)wrongClassified));
}
}