static public Pj CalculatePj(Sij sij, int time)//计算Pj
{
Pj pj = new Pj(time);
foreach (Sentence sentence in Variable.Sentences)
{
foreach (Labelset labelset in sij.Value[sentence].Keys)
{
if (pj.Value.ContainsKey(labelset))
{
pj.Value[labelset] += sij.Value[sentence][labelset];
}
else
{
pj.Value.Add(labelset, sij.Value[sentence][labelset]);
}
}
}
if (Variable.PjDividSentenceCount)
{
foreach (Labelset labelset in pj.Value.Keys.ToArray())
{
pj.Value[labelset] /= Variable.Sentences.Count;
}
}
//Variable.OutputFile.WriteLine(pj.ToString(DependentVariable.NumberOfIntlabel));
return(pj);//对于所有j，pj的和为1
}
//初始化
static public Sij InitializeSij(Label[] labels, int groupIndex)
{
Sij sij = new Sij(1);
foreach (Sentence sentence in Variable.Sentences)
{
sij.Value.Add(sentence, new Dictionary <Labelset, double>());
//Sij: Dictionary<句子，Dictionary<标签，次数>> 句子i被标为j的次数
foreach (Annotation annotation in sentence.AnnotaitonGroups[groupIndex].AnnotatorAnnotationDic.Values)
{
Labelset labelset = annotation.ToLabelset(labels);
if (sij.Value[sentence].ContainsKey(labelset))
{
++sij.Value[sentence][labelset];
}
else
{
sij.Value[sentence].Add(labelset, 1);
}
}
//Sij: Dictionary<句子，Dictionary<标签，概率>> 句子i被标为j的概率
foreach (Labelset labelset in sij.Value[sentence].Keys.ToArray())
{
sij.Value[sentence][labelset] /= Variable.NumberOfAnnotationsPerSentenceAfterGrouping;
}
}
return(sij);
}
static public void PriorPj(ref Pj pj, ref Mcj mcj, Sij sij, int time)
{
if (Variable.PriorP.Contains(PriorP.Pj) || Variable.PriorP.Contains(PriorP.ConditionalPj))
{
pj = CoreFunction.CalculatePj(sij, time);
}
if (Variable.PriorP.Contains(PriorP.Mcj) || Variable.PriorP.Contains(PriorP.ConditionalMcj))
{
mcj = ConsistencyFunction.CalculateMcj(sij, time);
}
}
static public void Intgerate(Label[] labels, int groupIndex, ref Sij sij)
{
Pj pj = null;
Pdata pdata = null;
Mcj mcj = null;
IDictionary <Tuple <Labelset, Labelset>, double> labelsetPairFrequencyForPj = null;//-1(开头)不能用来索引数组，故只能使用哈希表
IDictionary <Tuple <Character, Character>, IDictionary <Tuple <Labelset, Labelset>, double> > labelsetPairFrequencyForMcj = null;
IDictionary <Tuple <Sentence, Sentence>, IDictionary <Tuple <Labelset, Labelset>, double> > labelsetPairFrequencyForSij = null;
IDictionary <Label, IDictionary <Tuple <Labelset, Labelset>, double> > independentLabelsetPairFrequencyForPj = null;
IDictionary <Tuple <Character, Character>, IDictionary <Label, IDictionary <Tuple <Labelset, Labelset>, double> > > independentLabelsetPairFrequencyForMcj = null;
IDictionary <Tuple <Sentence, Sentence>, IDictionary <Label, IDictionary <Tuple <Labelset, Labelset>, double> > > independentLabelsetPairFrequencyForSij = null;
if (Variable.PriorP.Contains(PriorP.ConditionalPj))//计算在一篇（一次）标注中的转移频率
{
ChoiceFunction.InitializationOfLabelsetPairFrequencyForPj(labels, groupIndex, ref labelsetPairFrequencyForPj, ref independentLabelsetPairFrequencyForPj);
}
if (Variable.PriorP.Contains(PriorP.ConditionalMcj))
{
ChoiceFunction.InitializationOfLabelsetPairFrequencyForMcj(labels, groupIndex, ref labelsetPairFrequencyForMcj, ref independentLabelsetPairFrequencyForMcj);
}
if (Variable.PriorP.Contains(PriorP.ConditionalSij))
{
ChoiceFunction.InitializationOfLabelsetPairFrequencyForSij(labels, groupIndex, ref labelsetPairFrequencyForSij, ref independentLabelsetPairFrequencyForSij);
}
IList <double> Pdatas = new List <double>();
for (int time = 1; time <= Variable.ConvergeTimeThreshold; ++time)
{
//计算Pj，mcj（Consistency：角色c有j标签的概率）
ChoiceFunction.PriorPj(ref pj, ref mcj, sij, time);
//计算π
PAkjl pakjl = CalculatePAkjl(labels, sij, time, groupIndex);
//计算Sij
if (CalculatePdataAndSij(labels, ref sij, pj, pakjl, mcj, ref pdata, groupIndex, Pdatas, labelsetPairFrequencyForPj, labelsetPairFrequencyForMcj, labelsetPairFrequencyForSij))
{
break;
}
if (Variable.PriorP.Contains(PriorP.ConditionalPj))
{
ChoiceFunction.UpdateLabelsetPairFrequencyForPj(sij, ref labelsetPairFrequencyForPj, labels, ref independentLabelsetPairFrequencyForPj);
}
if (Variable.PriorP.Contains(PriorP.ConditionalMcj))
{
ChoiceFunction.UpdateLabelsetPairFrequencyForMcj(sij, ref labelsetPairFrequencyForMcj, labels, ref independentLabelsetPairFrequencyForMcj);
}
}
}
//包括独立与非独立
static public void UpdateLabelsetPairFrequencyForMcj(Sij sij, ref IDictionary <Tuple <Character, Character>, IDictionary <Tuple <Labelset, Labelset>, double> > LabelsetPairFrequencyForCharacter, Label[] labels, ref IDictionary <Tuple <Character, Character>, IDictionary <Label, IDictionary <Tuple <Labelset, Labelset>, double> > > independentLabelsetPairFrequencyForCharacter)
{
switch (Variable.Relation)
{
case RelationScheme.RenewLower:
#region 取相邻两者最小值作为联合概率
LabelsetPairFrequencyForCharacter = RelationFunction.RenewLabelsetPairFrequencyForCharacter(false, sij);
#endregion
break;
case RelationScheme.UpdateLower:
#region 取相邻两者最小值作为联合概率，不重新计算
RelationFunction.UpdateLabelsetPairFrequencyForCharacter(false, sij, ref LabelsetPairFrequencyForCharacter);
#endregion
break;
case RelationScheme.RenewOne:
#region 根据新结果重新计算labelPair的频率
LabelsetPairFrequencyForCharacter = RelationFunction.RenewLabelsetPairFrequencyForCharacter(true, sij);
#endregion
break;
case RelationScheme.UpdateOne:
#region 根据新结果将labelPair的频率添加到原有频率中，不重新计算
RelationFunction.UpdateLabelsetPairFrequencyForCharacter(true, sij, ref LabelsetPairFrequencyForCharacter);
#endregion
break;
case RelationScheme.AllLower:
LabelsetPairFrequencyForCharacter = RelationFunction.AllLabelsetPairFrequencyForCharacter(sij);
break;
case RelationScheme.IndependentRenewLower:
independentLabelsetPairFrequencyForCharacter = RelationFunction.RenewIndependentLabelPairFreuquencyForCharacter(false, labels, sij);
break;
case RelationScheme.IndependentRenewOne:
independentLabelsetPairFrequencyForCharacter = RelationFunction.RenewIndependentLabelPairFreuquencyForCharacter(true, labels, sij);
break;
}
}
public Sij(Sij sij)//复制初始化值，用于PjOrPij.PijNotChange时
{
Value = sij.Value;
Time = sij.Time;
}
static public PAkjl CalculatePAkjl(Label[] labels, Sij sij, int time, int groupIndex)//计算π
{
PAkjl pakjl = new PAkjl(time);
IDictionary <Annotator, IDictionary <Labelset, double> > denominator = new Dictionary <Annotator, IDictionary <Labelset, double> >();//Dictionary<人，Dictionary<标签j，值>>：𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠𝑒𝑛𝑡𝑒𝑛𝑐𝑒𝑠 𝑎𝑛𝑛𝑜𝑡𝑎𝑡𝑜𝑟 k 𝑟𝑒𝑐𝑜𝑟𝑑𝑠 𝑤ℎ𝑒𝑛 j 𝑖𝑠 𝑐𝑜𝑟𝑟𝑒𝑐𝑡，分母
foreach (Annotator annotator in GroupVariable.AnnotatorGroups[groupIndex])
{
pakjl.Value.Add(annotator, new Dictionary <Labelset, IDictionary <Labelset, double> >());//π本身充当分子
denominator.Add(annotator, new Dictionary <Labelset, double>());
}
//计算分子分母
foreach (Sentence sentence in Variable.Sentences)
{
foreach (Annotator annotator in sentence.AnnotaitonGroups[groupIndex].AnnotatorAnnotationDic.Keys) //只考虑这个人对于这句的标注，不考虑这个人对其他句的标注（如果这个人不在这个其他句的本组数据里）
{
foreach (Labelset labelsetj in sij.Value[sentence].Keys) //正确标签
{
//计算分子
if (!pakjl.Value[annotator].ContainsKey(labelsetj))
{
pakjl.Value[annotator].Add(labelsetj, new Dictionary <Labelset, double>());
}
Labelset labelsetl = sentence.AnnotaitonGroups[groupIndex].AnnotatorAnnotationDic[annotator].ToLabelset(labels);
if (pakjl.Value[annotator][labelsetj].ContainsKey(labelsetl))
{
pakjl.Value[annotator][labelsetj][labelsetl] += sij.Value[sentence][labelsetj];
}
else
{
pakjl.Value[annotator][labelsetj].Add(labelsetl, sij.Value[sentence][labelsetj]);
}
//计算分母
if (denominator[annotator].ContainsKey(labelsetj))
{
denominator[annotator][labelsetj] += sij.Value[sentence][labelsetj];
}
else
{
denominator[annotator].Add(labelsetj, sij.Value[sentence][labelsetj]);
}
}
}
}
//计算π
//要平滑：1.如果a > b，则让a/a > b/b（如不平滑，则都为一）；2.求sij时，多个π相乘会浮点溢出，使sij等于0
//此处是改进，只平滑π里有的，不平滑π里未出现过的jl对。经验证，Masatyan会改善
IDictionary <Smoothing, double[]> smoothingNumber = Function.SmoothingNumber(Math.Pow(2, labels.Length));
foreach (Annotator annotator in pakjl.Value.Keys)//人
{
foreach (Labelset labelsetj in pakjl.Value[annotator].Keys)
{
foreach (Labelset labelsetl in pakjl.Value[annotator][labelsetj].Keys.ToArray())
{
if (Variable.SmoothPajl != Smoothing.None)
{
if (denominator[annotator][labelsetj] != 0)
{
pakjl.Value[annotator][labelsetj][labelsetl] = (pakjl.Value[annotator][labelsetj][labelsetl] + smoothingNumber[Variable.SmoothPajl][0]) / (denominator[annotator][labelsetj] + smoothingNumber[Variable.SmoothPajl][1]);
}
else
{
pakjl.Value[annotator][labelsetj][labelsetj] = 1;//Dic赋值时，没有会直接添进去
}
}
else
{
pakjl.Value[annotator][labelsetj][labelsetl] /= denominator[annotator][labelsetj];
}
//if (denominator[annotator][labelsetj] != 0)
// pakjl.Value[annotator][labelsetj][labelsetl] /= denominator[annotator][labelsetj];
//else
// pakjl.Value[annotator][labelsetj][labelsetl] = 1 / Math.Pow(2, labels.Length);
}
}
}
return(pakjl);
}
//计算Sij和Pdata
static public bool CalculatePdataAndSij(Label[] labels, ref Sij sij, Pj pj, PAkjl pakjl, Mcj mcj, ref Pdata pdata, int groupIndex, IList <double> pdatas,
IDictionary <Tuple <Labelset, Labelset>, double> labelsetPairFrequencyForPj, IDictionary <Tuple <Character, Character>, IDictionary <Tuple <Labelset, Labelset>, double> > labelsetPairFrequencyForMcj,
IDictionary <Tuple <Sentence, Sentence>, IDictionary <Tuple <Labelset, Labelset>, double> > labelsetPairFrequencyForSij)
{
bool isFinished = false;
IDictionary <Tuple <Labelset, Labelset>, double> conditionalPj = null;//转移概率
IDictionary <Tuple <Character, Character>, IDictionary <Tuple <Labelset, Labelset>, double> > conditionalMcj = null;
if (Variable.PriorP.Contains(PriorP.ConditionalPj))
{
conditionalPj = RelationFunction.CalculateConditionalPj(pj, labelsetPairFrequencyForPj);
}
if (Variable.PriorP.Contains(PriorP.ConditionalMcj))
{
conditionalMcj = RelationFunction.CalculateConditionalMcj(mcj, labelsetPairFrequencyForMcj);
}
//sij的分子
IDictionary <Sentence, IDictionary <Labelset, double> > numerator = new Dictionary <Sentence, IDictionary <Labelset, double> >();
//sij的分母（P(data on i)）
IDictionary <Sentence, double> denominator = new Dictionary <Sentence, double>();
//计算分子
foreach (Sentence sentence in Variable.Sentences)
{
numerator.Add(sentence, new Dictionary <Labelset, double>());
#region 寻找需要遍历的j
//应该用pj中的j，即当前分组中出现过的所有标注情况，不能用sij现有的。因为要用pajl和pj重新计算sij，已与sij现有值无关，虽然对Boku来说结果一样
ICollection <Labelset> labelsets = null;
if (Variable.PriorP.Contains(PriorP.Sij) || Variable.PriorP.Contains(PriorP.ConditionalSij))
{
labelsets = sij.Value[sentence].Keys;
}
else if (Variable.PriorP.Contains(PriorP.Mcj) || Variable.PriorP.Contains(PriorP.ConditionalMcj))
{
labelsets = mcj.Value[sentence.Character].Keys;
}
else if (Variable.PriorP.Contains(PriorP.Pj) || Variable.PriorP.Contains(PriorP.ConditionalPj))
{
labelsets = pj.Value.Keys;//pj里只包含所有句出现过的所有标注情况，所以最多遍历pj即可，不需要for(int j = 0; j < Math.Pow(2, labels.Length); ++j)
}
else//全部的
{
labelsets = new List <Labelset>();
for (int j = 0; j < Math.Pow(2, labels.Length); ++j)
{
labelsets.Add(new Labelset(labels, j));
}
}
#endregion
//开始计算
foreach (Labelset labelsetj in labelsets)//此时结果会好一些（masa: group3, group5）
{
double valueOfNumerator = 1;
#region（公式 (5)）
foreach (Annotator annotator in sentence.AnnotaitonGroups[groupIndex].AnnotatorAnnotationDic.Keys)
{
Labelset labelsetl = sentence.AnnotaitonGroups[groupIndex].AnnotatorAnnotationDic[annotator].ToLabelset(labels);
if (pakjl.Value[annotator].ContainsKey(labelsetj))
{
if (pakjl.Value[annotator][labelsetj].ContainsKey(labelsetl))
{
valueOfNumerator *= pakjl.Value[annotator][labelsetj][labelsetl];
}
else
{
valueOfNumerator = 0;
break;
}
}
//else valueOfNumerator *= 1 / Math.Pow(2, labels.Length);//如果annotator没标过正确为j的句子，则认为此annotator对此j来说，所有可能标的标签l概率相等（对masa有用，没有的话（相当于乘以1）结果很差；boku进不来）
else
{
valueOfNumerator = 0; break;
} //相当于valueOfNumerator*0；boku进不来；此时结果会好一些（masa: group3）
}
if (valueOfNumerator == 0)
{
continue;
}
#endregion
#region 公式(5)*(6)
foreach (PriorP p in Variable.PriorP)
{
switch (p)
{
case PriorP.Pj:
valueOfNumerator *= pj.Value[labelsetj];
break;
case PriorP.Mcj:
valueOfNumerator *= mcj.Value[sentence.Character][labelsetj]; //Consistency的关键
break;
case PriorP.Sij:
valueOfNumerator *= sij.Value[sentence][labelsetj];
break;
case PriorP.ConditionalPj:
{
if (sentence.ID != 0)
{
bool finded = false;
double optimalPreLabelsetValue = 0;
double conditonalPj = 0;
int n = 1;
foreach (KeyValuePair <Labelset, double> labelsetPre in sij.SortLabelsets(Variable.Sentences[sentence.ID - 1])) //找出前一句最优标注和其概率
{
Tuple <Labelset, Labelset> labelsetPair = Tuple.Create(labelsetPre.Key, labelsetj);
if (!finded)
{
if (conditionalPj.ContainsKey(labelsetPair))
{
finded = true;
conditonalPj = conditionalPj[labelsetPair];//找出前一句最优标注到这一句的转移概率
optimalPreLabelsetValue = labelsetPre.Value;
}
continue;
}
if (finded)
{
if (labelsetPre.Value == optimalPreLabelsetValue && conditionalPj.ContainsKey(labelsetPair)) //最优标注可能不只一个（概率相同，同为最大），随意要继续遍历
{
conditonalPj += conditionalPj[labelsetPair];
++n;
}
else
{
break;
}
}
}
valueOfNumerator *= conditonalPj / n; //n：最优标注的个数
}
else
{
Tuple <Labelset, Labelset> labelsetPair = Tuple.Create(new Labelset(true), labelsetj);
if (conditionalPj.ContainsKey(labelsetPair))
{
valueOfNumerator *= conditionalPj[labelsetPair];
}
else
{
valueOfNumerator = 0;
}
}
}
break;
#region Mcj
case PriorP.ConditionalMcj:
{
if (sentence.ID != 0)
{
bool finded = false;
double optimalPreLabelsetValue = 0;
double maxConditonalMcj = 0;
int n = 1;
Sentence sentencePre = Variable.Sentences[sentence.ID - 1];
Tuple <Character, Character> characterPair = Tuple.Create(sentencePre.Character, sentence.Character);
foreach (KeyValuePair <Labelset, double> labelsetPre in sij.SortLabelsets(sentencePre))
{
Tuple <Labelset, Labelset> labelsetPair = Tuple.Create(labelsetPre.Key, labelsetj);
if (!finded)
{
if (conditionalMcj[characterPair].ContainsKey(labelsetPair))
{
finded = true;
maxConditonalMcj = conditionalMcj[characterPair][labelsetPair];
optimalPreLabelsetValue = labelsetPre.Value;
}
continue;
}
if (finded)
{
if (labelsetPre.Value == optimalPreLabelsetValue && conditionalMcj[characterPair].ContainsKey(labelsetPair))
{
maxConditonalMcj += conditionalMcj[characterPair][labelsetPair];
++n;
}
else
{
break;
}
}
}
valueOfNumerator *= maxConditonalMcj / n;
}
else
{
Tuple <Character, Character> characterPair = Tuple.Create(new Character("##"), sentence.Character);
Tuple <Labelset, Labelset> labelsetPair = Tuple.Create(new Labelset(true), labelsetj);
if (conditionalMcj[characterPair].ContainsKey(labelsetPair))
{
valueOfNumerator *= conditionalMcj[characterPair][labelsetPair];
}
else
{
valueOfNumerator = 0;
}
}
}
break;
#endregion
case PriorP.ConditionalSij:
{
if (sentence.ID != 0)
{
bool finded = false;
double optimalPreLabelsetValue = 0;
double conditonalPj = 0;
int n = 1;
Sentence sentencePre = Variable.Sentences[sentence.ID - 1];
Tuple <Sentence, Sentence> sentencePair = Tuple.Create(sentencePre, sentence);
foreach (KeyValuePair <Labelset, double> labelsetPre in sij.SortLabelsets(sentencePre))
{
Tuple <Labelset, Labelset> labelsetPair = Tuple.Create(labelsetPre.Key, labelsetj);
if (!finded)
{
if (labelsetPairFrequencyForSij[sentencePair].ContainsKey(labelsetPair))
{
finded = true;
conditonalPj = labelsetPairFrequencyForSij[sentencePair][labelsetPair] / labelsetPre.Value;
optimalPreLabelsetValue = labelsetPre.Value;
}
continue;
}
if (finded)
{
if (labelsetPre.Value == optimalPreLabelsetValue && labelsetPairFrequencyForSij[sentencePair].ContainsKey(labelsetPair))
{
conditonalPj += labelsetPairFrequencyForSij[sentencePair][labelsetPair] / labelsetPre.Value;
++n;
}
else
{
break;
}
}
}
valueOfNumerator *= conditonalPj / n;
}
else
{
Tuple <Sentence, Sentence> sentencePair = Tuple.Create(new Sentence(-1, "##"), sentence);
Tuple <Labelset, Labelset> labelsetPair = Tuple.Create(new Labelset(true), labelsetj);
if (labelsetPairFrequencyForSij[sentencePair].ContainsKey(labelsetPair))
{
valueOfNumerator *= labelsetPairFrequencyForSij[sentencePair][labelsetPair];
}
else
{
valueOfNumerator = 0;
}
}
}
break;
}
}
#endregion
if (valueOfNumerator != 0)
{
numerator[sentence].Add(labelsetj, valueOfNumerator);
}
}
#region 计算分母 (公式(7))
double valueOfDenominator = 0;
foreach (Labelset Labelsetq in numerator[sentence].Keys)//因为是加，故只需遍历numerator里有的标注，不需遍历所有标注
{
valueOfDenominator += numerator[sentence][Labelsetq];
}
denominator.Add(sentence, valueOfDenominator);
#endregion
}
//计算Pdata和Sij
pdata = pdata != null ? new Pdata(++pdata.Time, pdata.Value) : new Pdata(1, 0);
sij = new Sij(++sij.Time);
foreach (Sentence sentence in Variable.Sentences)
{
sij.Value.Add(sentence, new Dictionary <Labelset, double>());
foreach (Labelset labelset in numerator[sentence].Keys)
{
if (Variable.SijDividPDataOnI) //常规方法
{
sij.Value[sentence][labelset] = numerator[sentence][labelset] / denominator[sentence]; //Dic赋值时没有的元素会自动加
}
else
{
sij.Value[sentence][labelset] = numerator[sentence][labelset];
}
}
pdata.Value += -Math.Log(denominator[sentence]);
}
if (pdatas.Contains(pdata.Value) || (Math.Abs(pdata.MondifiedValue) <= Variable.ConvergeValueThreshold))
{
isFinished = true;
}
else
{
pdatas.Add(pdata.Value);
}
if (Variable.OutputPdata)
{
Variable.OutputFile.WriteLine(pdata.ToString());
}
return(isFinished);
}
