public bool LoadData(string[] inputfiles, out LdaDocument[] qaDocuments)
{
StreamReader sr = null;
int count = 0;
List<LdaDocument> documents = new List<LdaDocument>();
foreach (string inputfile in inputfiles)
{
Console.WriteLine("[Begin] read {0}", inputfile);
try
{
sr = new StreamReader(inputfile, Encoding.UTF8);
string queLine = null;
while ((queLine = sr.ReadLine()) != null)
{
if (count++ < 50000)
continue;
string[] tmp = queLine.Split(new string[] { "\t" }, StringSplitOptions.RemoveEmptyEntries);
if (tmp.Length < 2)
continue;
LdaDocument document = new LdaDocument();
List<int> queWordList = new List<int>();
var questionWords = tmp[0].Split(new char[] { '|' },StringSplitOptions.RemoveEmptyEntries);
foreach (var word in questionWords)
{
try
{
int index = int.Parse(word);
queWordList.Add(index);
}
catch (Exception e){ Console.WriteLine(e); }
}
List<int> quetags = new List<int>();
var tags = tmp[1].Split(new char[] { '|' }, StringSplitOptions.RemoveEmptyEntries);
foreach (var tag in tags)
{
quetags.Add(int.Parse(tag));
}
document.tag = quetags.ToArray();
document.doc = queWordList.ToArray();
documents.Add(document);
}
}
catch (Exception e) { Console.WriteLine(e); }
if (sr != null)
{
sr.Close();
}
Console.WriteLine("[End] read {0}", inputfile);
}
qaDocuments = documents.ToArray();
documents = null;
return true;
}
public GibbsLDA(LdaDocument[] documents, int vocbulrySize)
{
this.docs = documents;
this.docCount = documents.Length;
this.vocbulrySize = vocbulrySize;
}
public double[] GetRankDocHMMTrans(LdaDocument rankDoc)
{
if (rankDoc.doc == null || rankDoc.doc.Length == 0) { return new double[topicSum_K]; }
if (topicDis == null) { InitTopicDis(); }
if (phi == null) { ComputePhi(); }
if (topicToTopic == null) { ComputeTopicToTopic(); }
int queryLength = rankDoc.doc.Length;
rankDoc.wordTopic = new int[queryLength];
double[,] dis = new double[queryLength, topicSum_K];
int[,] preIndex = new int[queryLength, topicSum_K];
for (int i = 0; i < topicSum_K; i++)
{
double maxDis = -1;
for (int j = 0; j < topicSum_K; j++)
{
double curDis = topicDis[j] * topicToTopic[j][i];
if (curDis > maxDis)
{
maxDis = curDis;
}
}
dis[0, i] = maxDis * phi[i, rankDoc.doc[0]];
preIndex[0, i] = -1;
}
for (int i = 1; i < queryLength; i++)
{
for (int j = 0; j < topicSum_K; j++)
{
double maxDis = -1;
int index = -1;
for (int s = 0; s < topicSum_K; s++)
{
double curDis = dis[i - 1, s] * topicToTopic[s][j];
if (curDis > maxDis)
{
maxDis = curDis;
index = s;
}
}
dis[i, j] = maxDis * phi[j, rankDoc.doc[i]];
preIndex[i, j] = index;
}
}
int max = 0;
for (int i = 0; i < topicSum_K; i++)
{
if (dis[queryLength - 1, i] > dis[queryLength - 1, max]) { max = i; }
}
int[] rankTopicSum = new int[topicSum_K];
for (int i = queryLength - 1; i >= 0; i--)
{
rankDoc.wordTopic[i] = max;
rankTopicSum[max]++;
max = preIndex[i, max];
}
double[] rankTheta = new double[topicSum_K];
for (int i = 0; i < topicSum_K; i++) { rankTheta[i] = (rankTopicSum[i] + alpha) / (rankDoc.doc.Length + topicSum_K * alpha); }
return rankTheta;
}
public double[] GetRankDocumentTheta(LdaDocument rankDoc)
{
if (rankDoc.doc == null) { return null; }
double[] rankDocTheta = new double[topicSum_K];
# region InitRankDocumentState
int[] rankDocTopicSum = new int[topicSum_K];
int length = rankDoc.doc.Length;
int[,] rankDocWordTopicSum = new int[length, topicSum_K];
int[] wordTopic = new int[length];
for (int i = 0; i < length; i++)
{
int topic = random.Next(topicSum_K);
wordTopic[i] = topic;
rankDocTopicSum[topic]++;
rankDocWordTopicSum[i, topic]++;
}
rankDoc.wordTopic = wordTopic;
int rankNumstats = 0;
#endregion InitRankDocumentState
for (int it = 0; it < ITERATIONS; it++)
{
#region RankDocSample
if (rankDoc.wordTopic != null)
{
double[] p = new double[topicSum_K];
for (int i = 0; i < rankDoc.doc.Length; i++)
{
int word = rankDoc.doc[i];
int topic = rankDoc.wordTopic[i];
rankDocTopicSum[topic]--;
rankDocWordTopicSum[i, topic]--;
for (int j = 0; j < topicSum_K; j++)
{
p[j] = (wordTopicSum[word, j] + rankDocWordTopicSum[i, j] + beta)
/ (topicWordSum[j] + rankDocTopicSum[j] + vocbulrySize * beta) * (rankDocTopicSum[j] + alpha);
}
for (int k = 1; k < topicSum_K; k++) { p[k] += p[k - 1]; }
double u = random.NextDouble() * p[topicSum_K - 1];
for (topic = 0; topic < topicSum_K; topic++)
{
if (p[topic] > u) { break; }
}
rankDocWordTopicSum[i, topic]++;
rankDocTopicSum[topic]++;
rankDoc.wordTopic[i] = topic;
}
}
#endregion RankDocSample
if ((it > BURN_IN) && (SAMPLE_LAG > 0) && (it % SAMPLE_LAG == 0))
{
#region UpdateRankDocParams
for (int topic = 0; topic < topicSum_K; topic++)
{
rankDocTheta[topic] += ((rankDocTopicSum[topic] + alpha) / (rankDoc.doc.Length + topicSum_K * alpha));
}
rankNumstats++;
#endregion UpdateRankDocParams
}
}
#region ComputRankDocTheta
if (SAMPLE_LAG > 0)
{
for (int topic = 0; topic < topicSum_K; topic++)
{
rankDocTheta[topic] = rankDocTheta[topic] / rankNumstats;
}
}
else
{
for (int topic = 0; topic < topicSum_K; topic++)
{
rankDocTheta[topic] = ((rankDocTopicSum[topic] + alpha) / (rankDoc.doc.Length + topicSum_K * alpha));
}
}
#endregion ComputRankDocTheta
Console.Write(".");
return rankDocTheta;
}
public double TopicLikelihood(LdaDocument question, LdaDocument candidate)
{
var questionTheta = GetRankDocumentTheta(question);
var candidateTheta = GetRankDocumentTheta(candidate);
double score = 0;
for (int i = 0; i < topicSum_K; i++)
{
score += (questionTheta[i] * candidateTheta[i] / topicWordSum[i]);
}
return score * question.doc.Length;
}
