public virtual ICollection <string> FeaturesCnC(PaddedList <IN> cInfo, int loc)
{
ICollection <string> features = new List <string>();
CoreLabel c = cInfo[loc];
CoreLabel c1 = cInfo[loc + 1];
CoreLabel p = cInfo[loc - 1];
string charc = c.Get(typeof(CoreAnnotations.CharAnnotation));
string charc1 = c1.Get(typeof(CoreAnnotations.CharAnnotation));
string charp = p.Get(typeof(CoreAnnotations.CharAnnotation));
if (flags.useWordn)
{
features.Add(charc + "c");
features.Add(charc1 + "c1");
features.Add(charp + "p");
features.Add(charp + charc + "pc");
if (flags.useAs || flags.useMsr || flags.usePk || flags.useHk)
{
features.Add(charc + charc1 + "cc1");
features.Add(charp + charc1 + "pc1");
}
features.Add("|wordn");
}
return(features);
}
/// <summary>Extracts all the features from the input data at a certain index.</summary>
/// <param name="cInfo">The complete data set as a List of WordInfo</param>
/// <param name="loc">The index at which to extract features.</param>
public override ICollection <string> GetCliqueFeatures(PaddedList <IN> cInfo, int loc, Clique clique)
{
ICollection <string> features = Generics.NewHashSet();
if (clique == cliqueC)
{
AddAllInterningAndSuffixing(features, FeaturesC(cInfo, loc), "C");
}
else
{
if (clique == cliqueCpC)
{
AddAllInterningAndSuffixing(features, FeaturesCpC(cInfo, loc), "CpC");
AddAllInterningAndSuffixing(features, FeaturesCnC(cInfo, loc - 1), "CnC");
}
}
// else if (clique == cliqueCpCp2C) {
// addAllInterningAndSuffixing(features, featuresCpCp2C(cInfo, loc), "CpCp2C");
// } else if (clique == cliqueCpCp2Cp3C) {
// addAllInterningAndSuffixing(features, featuresCpCp2Cp3C(cInfo, loc), "CpCp2Cp3C");
// } else if (clique == cliqueCpCp2Cp3Cp4C) {
// addAllInterningAndSuffixing(features, featuresCpCp2Cp3Cp4C(cInfo, loc), "CpCp2Cp3Cp4C");
// } else if (clique == cliqueCpCp2Cp3Cp4Cp5C) {
// addAllInterningAndSuffixing(features, featuresCpCp2Cp3Cp4Cp5C(cInfo, loc), "CpCp2Cp3Cp4Cp5C");
// }
return(features);
}
// end featuresCpC
/// <summary>
/// For a CRF, this shouldn't be necessary, since the features duplicate
/// those from CpC, but Huihsin found some valuable, presumably becuase
/// it modified the regularization a bit.
/// </summary>
/// <param name="cInfo">The list of characters</param>
/// <param name="loc">Position of c in list</param>
/// <returns>Collection of String features (sparse set of boolean features</returns>
protected internal virtual ICollection <string> FeaturesCnC <_T0>(PaddedList <_T0> cInfo, int loc)
where _T0 : CoreLabel
{
ICollection <string> features = new List <string>();
if (flags.useWordn)
{
CoreLabel c = cInfo[loc];
CoreLabel c2 = cInfo[loc + 1];
CoreLabel p = cInfo[loc - 1];
CoreLabel p2 = cInfo[loc - 2];
string charc = c.GetString <CoreAnnotations.CharAnnotation>();
string charc2 = c2.GetString <CoreAnnotations.CharAnnotation>();
string charp = p.GetString <CoreAnnotations.CharAnnotation>();
string charp2 = p2.GetString <CoreAnnotations.CharAnnotation>();
features.Add(charc + "c");
features.Add(charc2 + "c2");
features.Add(charp + "p");
features.Add(charp2 + "p2");
features.Add(charp2 + charp + "p2p");
features.Add(charp + charc + "pc");
features.Add(charc + charc2 + "cc2");
features.Add(charp + "-" + charc2 + "pc2");
features.Add("cliqueCnC");
}
return(features);
}
protected internal virtual ICollection <string> FeaturesC(PaddedList <In> cInfo, int loc)
{
ICollection <string> features = new List <string>();
CoreLabel c = cInfo[loc];
CoreLabel n = cInfo[loc + 1];
CoreLabel n2 = cInfo[loc + 2];
CoreLabel p = cInfo[loc - 1];
CoreLabel p2 = cInfo[loc - 2];
string charc = c.Get(typeof(CoreAnnotations.CharAnnotation));
string charn = n.Get(typeof(CoreAnnotations.CharAnnotation));
string charn2 = n2.Get(typeof(CoreAnnotations.CharAnnotation));
string charp = p.Get(typeof(CoreAnnotations.CharAnnotation));
string charp2 = p2.Get(typeof(CoreAnnotations.CharAnnotation));
// Default feature set...a 5 character window
// plus a few other language-independent features
features.Add(charc + "-c");
features.Add(charn + "-n1");
features.Add(charn2 + "-n2");
features.Add(charp + "-p");
features.Add(charp2 + "-p2");
// Length feature
if (charc.Length > 1)
{
features.Add("length");
}
// Character-level class features
bool seenPunc = false;
bool seenDigit = false;
for (int i = 0; i < limit; ++i)
{
char charcC = charc[i];
seenPunc = seenPunc || Characters.IsPunctuation(charcC);
seenDigit = seenDigit || char.IsDigit(charcC);
string cuBlock = Characters.UnicodeBlockStringOf(charcC);
features.Add(cuBlock + "-uBlock");
string cuType = char.GetType(charcC).ToString();
features.Add(cuType + "-uType");
}
if (seenPunc)
{
features.Add("haspunc");
}
if (seenDigit)
{
features.Add("hasdigit");
}
// Token-level features
string word = c.Word();
int index = c.Index();
features.Add(Math.Min(MaxBefore, index) + "-before");
features.Add(Math.Min(MaxAfter, word.Length - charc.Length - index) + "-after");
features.Add(Math.Min(MaxLength, word.Length) + "-length");
// Indicator transition feature
features.Add("cliqueC");
return(features);
}
protected internal override ICollection <string> FeaturesC(PaddedList <In> cInfo, int loc)
{
ICollection <string> features = base.FeaturesC(cInfo, loc);
CoreLabel n3 = cInfo[loc + 3];
CoreLabel p3 = cInfo[loc - 3];
string charn3 = n3.Get(typeof(CoreAnnotations.CharAnnotation));
string charp3 = p3.Get(typeof(CoreAnnotations.CharAnnotation));
// a 7 character window instead of a 5 character window
features.Add(charn3 + "-n3");
features.Add(charp3 + "-p3");
return(features);
}
protected internal virtual ICollection <string> FeaturesCpC(PaddedList <In> cInfo, int loc)
{
ICollection <string> features = new List <string>();
CoreLabel c = cInfo[loc];
CoreLabel p = cInfo[loc - 1];
string charc = c.Get(typeof(CoreAnnotations.CharAnnotation));
string charp = p.Get(typeof(CoreAnnotations.CharAnnotation));
features.Add(charc + charp + "-cngram");
// Indicator transition feature
features.Add("cliqueCpC");
return(features);
}
public override ICollection <string> GetCliqueFeatures(PaddedList <CoreLabel> info, int position, Clique clique)
{
ICollection <string> features = new HashSet <string>();
foreach (CoreLabel l in info)
{
for (int i = 0; i < 10; i++)
{
features.Add("feat" + i + ":" + l.Word());
}
}
return(features);
}
//is EnglishPU
public virtual ICollection <string> FeaturesC(PaddedList <IN> cInfo, int loc)
{
ICollection <string> features = new List <string>();
CoreLabel c = cInfo[loc];
CoreLabel c1 = cInfo[loc + 1];
CoreLabel c2 = cInfo[loc + 2];
CoreLabel c3 = cInfo[loc + 3];
CoreLabel p = cInfo[loc - 1];
CoreLabel p2 = cInfo[loc - 2];
CoreLabel p3 = cInfo[loc - 3];
string charc = c.Get(typeof(CoreAnnotations.CharAnnotation));
string charc1 = c1.Get(typeof(CoreAnnotations.CharAnnotation));
string charc2 = c2.Get(typeof(CoreAnnotations.CharAnnotation));
string charc3 = c3.Get(typeof(CoreAnnotations.CharAnnotation));
string charp = p.Get(typeof(CoreAnnotations.CharAnnotation));
string charp2 = p2.Get(typeof(CoreAnnotations.CharAnnotation));
string charp3 = p3.Get(typeof(CoreAnnotations.CharAnnotation));
if (flags.useWord1)
{
// features.add(charc +"c");
// features.add(charc1+"c1");
// features.add(charp +"p");
// features.add(charp +charc +"pc");
// if(flags.useAs || flags.useMsr || flags.usePk || flags.useHk){ //msr, as
// features.add(charc +charc1 +"cc1");
// features.add(charp + charc1 +"pc1");
// }
features.Add(charc + "::c");
features.Add(charc1 + "::c1");
features.Add(charp + "::p");
features.Add(charp2 + "::p2");
// trying to restore the features that Huishin described in SIGHAN 2005 paper
features.Add(charc + charc1 + "::cn");
features.Add(charp + charc + "::pc");
features.Add(charp + charc1 + "::pn");
features.Add(charp2 + charp + "::p2p");
features.Add(charp2 + charc + "::p2c");
features.Add(charc2 + charc + "::n2c");
features.Add("|word1");
}
return(features);
}
/// <summary>Extracts all the features from the input data at a certain index.</summary>
/// <param name="cInfo">The complete data set as a List of WordInfo</param>
/// <param name="loc">The index at which to extract features.</param>
public override ICollection <string> GetCliqueFeatures(PaddedList <In> cInfo, int loc, Clique clique)
{
ICollection <string> features = Generics.NewHashSet();
if (clique == cliqueC)
{
AddAllInterningAndSuffixing(features, FeaturesC(cInfo, loc), "C");
}
else
{
if (clique == cliqueCpC)
{
AddAllInterningAndSuffixing(features, FeaturesCpC(cInfo, loc), "CpC");
}
else
{
if (clique == cliqueCp2C)
{
AddAllInterningAndSuffixing(features, FeaturesCp2C(cInfo, loc), "Cp2C");
}
else
{
if (clique == cliqueCp3C)
{
AddAllInterningAndSuffixing(features, FeaturesCp3C(cInfo, loc), "Cp3C");
}
}
}
}
string domain = cInfo[loc].Get(typeof(CoreAnnotations.DomainAnnotation));
if (domain != null)
{
ICollection <string> domainFeatures = Generics.NewHashSet();
foreach (string feature in features)
{
domainFeatures.Add(feature + DomainMarker + domain);
}
Sharpen.Collections.AddAll(features, domainFeatures);
}
return(features);
}
// end featuresCpCp2C
protected internal virtual ICollection <string> FeaturesCpCp2Cp3C <_T0>(PaddedList <_T0> cInfo, int loc)
where _T0 : CoreLabel
{
ICollection <string> features = new List <string>();
if (flags.use4Clique && flags.maxLeft >= 3)
{
CoreLabel c = cInfo[loc];
CoreLabel c2 = cInfo[loc + 1];
CoreLabel p = cInfo[loc - 1];
CoreLabel p2 = cInfo[loc - 2];
CoreLabel p3 = cInfo[loc - 3];
string charc = c.GetString <CoreAnnotations.CharAnnotation>();
string charp = p.GetString <CoreAnnotations.CharAnnotation>();
string charp2 = p2.GetString <CoreAnnotations.CharAnnotation>();
string charp3 = p3.GetString <CoreAnnotations.CharAnnotation>();
int cI = c.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypec = (cI != null ? cI.ToString() : string.Empty);
int c2I = c2.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypec2 = (c2I != null ? c2I.ToString() : string.Empty);
int pI = p.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypep = (pI != null ? pI.ToString() : string.Empty);
int p2I = p2.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypep2 = (p2I != null ? p2I.ToString() : string.Empty);
int p3I = p3.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypep3 = (p3I != null ? p3I.ToString() : string.Empty);
if (flags.useLongSequences)
{
features.Add(charp3 + charp2 + charp + charc + "p3p2pc");
}
if (flags.useUnicodeType4gram || flags.useUnicodeType5gram)
{
features.Add(uTypep3 + "-" + uTypep2 + "-" + uTypep + "-" + uTypec + "-uType4");
}
if (flags.useUnicodeType5gram)
{
features.Add(uTypep3 + "-" + uTypep2 + "-" + uTypep + "-" + uTypec + "-" + uTypec2 + "-uType5");
}
features.Add("cliqueCpCp2Cp3C");
}
return(features);
}
protected internal override ICollection <string> FeaturesCpC(PaddedList <IN> cInfo, int loc)
{
ICollection <string> features = base.FeaturesCpC(cInfo, loc);
CoreLabel c = cInfo[loc];
// "Wrapper" feature: identity of first and last two chars of the current word.
// This helps detect ma+_+sh in dialect, as well as avoiding segmenting possessive
// pronouns if the word starts with al-.
if (c.Word().Length > 3)
{
string start = Sharpen.Runtime.Substring(c.Word(), 0, 2);
string end = Sharpen.Runtime.Substring(c.Word(), c.Word().Length - 2);
if (c.Index() == 2)
{
features.Add(start + "_" + end + "-begin-wrap");
}
if (c.Index() == c.Word().Length - 1)
{
features.Add(start + "_" + end + "-end-wrap");
}
}
return(features);
}
public override CRFDatum <IList <string>, CRFLabel> MakeDatum(IList <IN> info, int loc, IList <FeatureFactory <IN> > featureFactories)
{
pad.Set(typeof(CoreAnnotations.AnswerAnnotation), flags.backgroundSymbol);
PaddedList <IN> pInfo = new PaddedList <IN>(info, pad);
IList <IList <string> > features = new List <IList <string> >();
ICollection <Clique> done = Generics.NewHashSet();
for (int i = 0; i < windowSize; i++)
{
IList <string> featuresC = new List <string>();
IList <Clique> windowCliques = FeatureFactory.GetCliques(i, 0);
windowCliques.RemoveAll(done);
Sharpen.Collections.AddAll(done, windowCliques);
foreach (Clique c in windowCliques)
{
foreach (FeatureFactory <IN> featureFactory in featureFactories)
{
Sharpen.Collections.AddAll(featuresC, featureFactory.GetCliqueFeatures(pInfo, loc, c));
}
}
if (testTime && i == 0)
{
// this feature is only present at test time and only appears
// in cliques of size 1 (i.e., cliques with window=0)
featuresC.Add(Bias);
}
features.Add(featuresC);
}
int[] labels = new int[windowSize];
for (int i_1 = 0; i_1 < windowSize; i_1++)
{
string answer = pInfo[loc + i_1 - windowSize + 1].Get(typeof(CoreAnnotations.AnswerAnnotation));
labels[i_1] = classIndex.IndexOf(answer);
}
return(new CRFDatum <IList <string>, CRFLabel>(features, new CRFLabel(labels), null));
}
/// <summary>Extracts all the features from the input data at a certain index.</summary>
/// <param name="cInfo">The complete data set as a List of WordInfo</param>
/// <param name="loc">The index at which to extract features.</param>
public override ICollection <string> GetCliqueFeatures(PaddedList <IN> cInfo, int loc, Clique clique)
{
ICollection <string> features = Generics.NewHashSet();
if (clique == cliqueC)
{
AddAllInterningAndSuffixing(features, FeaturesC(cInfo, loc), "C");
}
else
{
if (clique == cliqueCpC)
{
AddAllInterningAndSuffixing(features, FeaturesCpC(cInfo, loc), "CpC");
AddAllInterningAndSuffixing(features, FeaturesCnC(cInfo, loc - 1), "CnC");
}
else
{
if (clique == cliqueCpCp2C)
{
AddAllInterningAndSuffixing(features, FeaturesCpCp2C(cInfo, loc), "CpCp2C");
}
else
{
if (clique == cliqueCpCp2Cp3C)
{
AddAllInterningAndSuffixing(features, FeaturesCpCp2Cp3C(cInfo, loc), "CpCp2Cp3C");
}
}
}
}
if (Debug > 0)
{
EncodingPrintWriter.Err.Println("For " + cInfo[loc] + ", features: " + features, "UTF-8");
}
return(features);
}
/* (non-Javadoc)
* @see edu.stanford.nlp.sequences.FeatureFactory#getCliqueFeatures(edu.stanford.nlp.util.PaddedList, int, edu.stanford.nlp.sequences.Clique)
*/
public override ICollection GetCliqueFeatures(PaddedList info, int position, Clique clique)
{
// TODO Auto-generated method stub
return(null);
}
//end of CnC
/// <summary>Second order clique features</summary>
/// <param name="cInfo">The list of characters</param>
/// <param name="loc">Position of c in list</param>
/// <returns>Collection of String features (sparse set of boolean features</returns>
protected internal virtual ICollection <string> FeaturesCpCp2C <_T0>(PaddedList <_T0> cInfo, int loc)
where _T0 : CoreLabel
{
ICollection <string> features = new List <string>();
CoreLabel c = cInfo[loc];
CoreLabel c2 = cInfo[loc + 1];
CoreLabel c3 = cInfo[loc + 2];
CoreLabel p = cInfo[loc - 1];
CoreLabel p2 = cInfo[loc - 2];
CoreLabel p3 = cInfo[loc - 3];
string charc = c.GetString <CoreAnnotations.CharAnnotation>();
string charc2 = c2.GetString <CoreAnnotations.CharAnnotation>();
string charc3 = c3.GetString <CoreAnnotations.CharAnnotation>();
string charp = p.GetString <CoreAnnotations.CharAnnotation>();
string charp2 = p2.GetString <CoreAnnotations.CharAnnotation>();
string charp3 = p3.GetString <CoreAnnotations.CharAnnotation>();
// N-gram features. N is up to 3
if (flags.useWord3)
{
features.Add(charc + "::c");
features.Add(charc2 + "::n");
features.Add(charp + "::p");
features.Add(charp2 + "::p2");
// trying to restore the features that Huihsin described in SIGHAN 2005 paper
features.Add(charc + charc2 + "::cn");
features.Add(charc + charc2 + charc3 + "::cnn2");
features.Add(charp + charc + "::pc");
features.Add(charp + charc2 + "::pn");
features.Add(charp2 + charp + "::p2p");
features.Add(charp3 + charp2 + charp + "::p3p2p");
features.Add(charp2 + charc + "::p2c");
features.Add(charc + charc3 + "::cn2");
}
if (flags.useShapeStrings)
{
if (flags.useShapeStrings1)
{
features.Add(p.GetString <CoreAnnotations.ShapeAnnotation>() + "ps");
features.Add(c.GetString <CoreAnnotations.ShapeAnnotation>() + "cs");
features.Add(c2.GetString <CoreAnnotations.ShapeAnnotation>() + "c2s");
}
if (flags.useShapeStrings3)
{
features.Add(p.GetString <CoreAnnotations.ShapeAnnotation>() + c.GetString <CoreAnnotations.ShapeAnnotation>() + c2.GetString <CoreAnnotations.ShapeAnnotation>() + "pscsc2s");
}
if (flags.useShapeStrings4)
{
features.Add(p2.GetString <CoreAnnotations.ShapeAnnotation>() + p.GetString <CoreAnnotations.ShapeAnnotation>() + c.GetString <CoreAnnotations.ShapeAnnotation>() + c2.GetString <CoreAnnotations.ShapeAnnotation>() + "p2spscsc2s");
}
if (flags.useShapeStrings5)
{
features.Add(p2.GetString <CoreAnnotations.ShapeAnnotation>() + p.GetString <CoreAnnotations.ShapeAnnotation>() + c.GetString <CoreAnnotations.ShapeAnnotation>() + c2.GetString <CoreAnnotations.ShapeAnnotation>() + c3.GetString <CoreAnnotations.ShapeAnnotation
>() + "p2spscsc2sc3s");
}
if (flags.useWordShapeConjunctions2)
{
features.Add(p.GetString <CoreAnnotations.ShapeAnnotation>() + charc + "pscc");
features.Add(charp + c.GetString <CoreAnnotations.ShapeAnnotation>() + "pccs");
}
if (flags.useWordShapeConjunctions3)
{
features.Add(p2.GetString <CoreAnnotations.ShapeAnnotation>() + p.GetString <CoreAnnotations.ShapeAnnotation>() + charc + "p2spscc");
features.Add(p.GetString <CoreAnnotations.ShapeAnnotation>() + charc + c2.GetString <CoreAnnotations.ShapeAnnotation>() + "psccc2s");
features.Add(charc + c2.GetString <CoreAnnotations.ShapeAnnotation>() + c3.GetString <CoreAnnotations.ShapeAnnotation>() + "ccc2sc3s");
}
}
/*
* Radical N-gram features. N is upto 4.
* Smoothing method of N-gram, because there are too many characters in Chinese.
* (It works better than N-gram when they are used individually. less sparse)
*/
char rcharc;
char rcharc2;
char rcharp;
char rcharp2;
if (charc.Length == 0)
{
rcharc = 'n';
}
else
{
rcharc = RadicalMap.GetRadical(charc[0]);
}
if (charc2.Length == 0)
{
rcharc2 = 'n';
}
else
{
rcharc2 = RadicalMap.GetRadical(charc2[0]);
}
if (charp.Length == 0)
{
rcharp = 'n';
}
else
{
rcharp = RadicalMap.GetRadical(charp[0]);
}
if (charp2.Length == 0)
{
rcharp2 = 'n';
}
else
{
rcharp2 = RadicalMap.GetRadical(charp2[0]);
}
if (flags.useRad2)
{
features.Add(rcharc + "rc");
features.Add(rcharc2 + "rc2");
features.Add(rcharp + "rp");
features.Add(rcharp + rcharc + "rprc");
features.Add(rcharc + rcharc2 + "rcrc2");
features.Add(rcharp + rcharc + rcharc2 + "rprcrc2");
}
if (flags.useRad2b)
{
features.Add(rcharc + "rc");
features.Add(rcharc2 + "rc2");
features.Add(rcharp + "rp");
features.Add(rcharp + rcharc + "rprc");
features.Add(rcharc + rcharc2 + "rcrc2");
features.Add(rcharp2 + rcharp + "rp2rp");
}
features.Add("cliqueCpCp2C");
return(features);
}
// static methods
/// <summary>
/// This can be used to map from any IOB-style (i.e., "I-PERS" style labels)
/// or just categories representation to any other.
/// </summary>
/// <remarks>
/// This can be used to map from any IOB-style (i.e., "I-PERS" style labels)
/// or just categories representation to any other.
/// It can read and change any representation to other representations:
/// a 4 way representation of all entities, like S-PERS, B-PERS,
/// I-PERS, E-PERS for single word, beginning, internal, and end of entity
/// (IOBES or SBIEO); always marking the first word of an entity (IOB2 or BIO);
/// only marking specially the beginning of non-first
/// items of an entity sequences with B-PERS (IOB1);
/// the reverse IOE1 and IOE2; IO where everything is I-tagged; and
/// NOPREFIX, where no prefixes are written on category labels.
/// The last two representations are deficient in not allowing adjacent
/// entities of the same class to be represented, but nevertheless
/// convenient. Note that the background label is never given a prefix.
/// This code is very specific to the particular CoNLL way of labeling
/// classes for IOB-style encoding, but this notation is quite widespread.
/// It will work on any of these styles of input.
/// This will also recognize BILOU format (B=B, I=I, L=E, O=O, U=S).
/// It also works with lowercased names like i-org.
/// If the labels are not of the form "C-Y+", where C is a single character,
/// then they will be regarded as NOPREFIX labels.
/// This method updates the List tokens in place.
/// </remarks>
/// <param name="tokens">List of tokens (each a CoreLabel) in some style</param>
/// <param name="key">The key in the CoreLabel to change, commonly CoreAnnotations.AnswerAnnotation.class</param>
/// <param name="backgroundLabel">The background label, which gets special treatment</param>
/// <param name="style">Output style; one of iob[12], ioe[12], io, sbieo/iobes, noprefix</param>
/// <param name="intern">Whether to String-intern the new labels (may as well, small number!)</param>
public static void EntitySubclassify <Tok>(IList <TOK> tokens, Type key, string backgroundLabel, string style, bool intern)
where Tok : ICoreMap
{
int how;
string lowerStyle = style.ToLower(Locale.English);
switch (lowerStyle)
{
case "iob1":
{
how = 0;
break;
}
case "iob2":
case "bio":
{
how = 1;
break;
}
case "ioe1":
{
how = 2;
break;
}
case "ioe2":
{
how = 3;
break;
}
case "io":
{
how = 4;
break;
}
case "sbieo":
case "iobes":
{
how = 5;
break;
}
case "noprefix":
{
how = 6;
break;
}
case "bilou":
{
how = 7;
break;
}
default:
{
throw new ArgumentException("entitySubclassify: unknown style: " + style);
}
}
IList <TOK> paddedTokens = new PaddedList <TOK>(tokens, (TOK) new CoreLabel());
int size = paddedTokens.Count;
string[] newAnswers = new string[size];
for (int i = 0; i < size; i++)
{
TOK c = paddedTokens[i];
TOK p = paddedTokens[i - 1];
TOK n = paddedTokens[i + 1];
string cAns = c.Get(key);
string pAns = p.Get(key);
if (pAns == null)
{
pAns = backgroundLabel;
}
string nAns = n.Get(key);
if (nAns == null)
{
nAns = backgroundLabel;
}
string @base;
char prefix;
if (cAns.Length > 2 && cAns[1] == '-')
{
@base = Sharpen.Runtime.Substring(cAns, 2, cAns.Length);
prefix = char.ToUpperCase(cAns[0]);
}
else
{
@base = cAns;
prefix = ' ';
}
string pBase;
char pPrefix;
if (pAns.Length > 2 && pAns[1] == '-')
{
pBase = Sharpen.Runtime.Substring(pAns, 2, pAns.Length);
pPrefix = char.ToUpperCase(pAns[0]);
}
else
{
pBase = pAns;
pPrefix = ' ';
}
string nBase;
char nPrefix;
if (nAns.Length > 2 && nAns[1] == '-')
{
nBase = Sharpen.Runtime.Substring(nAns, 2, nAns.Length);
nPrefix = char.ToUpperCase(nAns[0]);
}
else
{
nBase = nAns;
nPrefix = ' ';
}
bool isStartAdjacentSame = IsSameEntityBoundary(pBase, pPrefix, @base, prefix);
bool isEndAdjacentSame = IsSameEntityBoundary(@base, prefix, nBase, nPrefix);
bool isFirst = IsDifferentEntityBoundary(pBase, @base) || isStartAdjacentSame;
bool isLast = IsDifferentEntityBoundary(@base, nBase) || isEndAdjacentSame;
string newAnswer = @base;
if ([email protected](backgroundLabel))
{
switch (how)
{
case 0:
{
// iob1, only B if adjacent
if (isStartAdjacentSame)
{
newAnswer = "B-" + @base;
}
else
{
newAnswer = "I-" + @base;
}
break;
}
case 1:
{
// iob2 always B at start
if (isFirst)
{
newAnswer = "B-" + @base;
}
else
{
newAnswer = "I-" + @base;
}
break;
}
case 2:
{
// ioe1
if (isEndAdjacentSame)
{
newAnswer = "E-" + @base;
}
else
{
newAnswer = "I-" + @base;
}
break;
}
case 3:
{
// ioe2
if (isLast)
{
newAnswer = "E-" + @base;
}
else
{
newAnswer = "I-" + @base;
}
break;
}
case 4:
{
newAnswer = "I-" + @base;
break;
}
case 5:
{
if (isFirst && isLast)
{
newAnswer = "S-" + @base;
}
else
{
if ((!isFirst) && isLast)
{
newAnswer = "E-" + @base;
}
else
{
if (isFirst && (!isLast))
{
newAnswer = "B-" + @base;
}
else
{
newAnswer = "I-" + @base;
}
}
}
break;
}
case 7:
{
// nothing to do on case 6 as it's just base
if (isFirst && isLast)
{
newAnswer = "U-" + @base;
}
else
{
if ((!isFirst) && isLast)
{
newAnswer = "L-" + @base;
}
else
{
if (isFirst && (!isLast))
{
newAnswer = "B-" + @base;
}
else
{
newAnswer = "I-" + @base;
}
}
}
break;
}
}
}
if (intern)
{
newAnswer = string.Intern(newAnswer);
}
newAnswers[i] = newAnswer;
}
for (int i_1 = 0; i_1 < size; i_1++)
{
TOK c = tokens[i_1];
c.Set(typeof(CoreAnnotations.AnswerAnnotation), newAnswers[i_1]);
}
}
protected internal virtual ICollection <string> FeaturesCpC <_T0>(PaddedList <_T0> cInfo, int loc)
where _T0 : CoreLabel
{
ICollection <string> features = new List <string>();
CoreLabel c = cInfo[loc];
CoreLabel c2 = cInfo[loc + 1];
CoreLabel c3 = cInfo[loc + 2];
CoreLabel p = cInfo[loc - 1];
CoreLabel p2 = cInfo[loc - 2];
CoreLabel p3 = cInfo[loc - 3];
string charc = c.GetString <CoreAnnotations.CharAnnotation>();
string charc2 = c2.GetString <CoreAnnotations.CharAnnotation>();
string charc3 = c3.GetString <CoreAnnotations.CharAnnotation>();
string charp = p.GetString <CoreAnnotations.CharAnnotation>();
string charp2 = p2.GetString <CoreAnnotations.CharAnnotation>();
string charp3 = p3.GetString <CoreAnnotations.CharAnnotation>();
int cI = c.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypec = (cI != null ? cI.ToString() : string.Empty);
int c2I = c2.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypec2 = (c2I != null ? c2I.ToString() : string.Empty);
int c3I = c3.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypec3 = (c3I != null ? c3I.ToString() : string.Empty);
int pI = p.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypep = (pI != null ? pI.ToString() : string.Empty);
int p2I = p2.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypep2 = (p2I != null ? p2I.ToString() : string.Empty);
if (flags.dictionary != null || flags.serializedDictionary != null)
{
DictionaryFeaturesCpC(typeof(CoreAnnotations.LBeginAnnotation), typeof(CoreAnnotations.LMiddleAnnotation), typeof(CoreAnnotations.LEndAnnotation), string.Empty, features, p2, p, c, c2);
}
if (flags.dictionary2 != null)
{
DictionaryFeaturesCpC(typeof(CoreAnnotations.D2_LBeginAnnotation), typeof(CoreAnnotations.D2_LMiddleAnnotation), typeof(CoreAnnotations.D2_LEndAnnotation), "-D2-", features, p2, p, c, c2);
}
/*
* N-gram features. N is upto 2.
*/
if (flags.useWord2)
{
// features.add(charc +"c");
// features.add(charc2+"c2");
// features.add(charp +"p");
// features.add(charp + charc +"pc");
// features.add(charc + charc2 +"cc2");
// // cdm: need hyphen so you can see which of charp or charc2 is null....
// features.add(charp + "-" + charc2 + "pc2");
features.Add(charc + "::c");
features.Add(charc2 + "::c1");
features.Add(charp + "::p");
features.Add(charp2 + "::p2");
// trying to restore the features that Huihsin described in SIGHAN 2005 paper
features.Add(charc + charc2 + "::cn");
// (*)
features.Add(charp + charc + "::pc");
features.Add(charp + charc2 + "::pn");
features.Add(charp2 + charp + "::p2p");
features.Add(charp2 + charc + "::p2c");
features.Add(charc2 + charc + "::n2c");
}
// todo: this is messed up: Same as one above at (*); should be cn2 = charc + charc3 + "::cn2"
if (flags.useFeaturesCpC4gram || flags.useFeaturesCpC5gram || flags.useFeaturesCpC6gram)
{
// todo: Both these features duplicate ones already in useWord2
features.Add(charp2 + charp + "p2p");
features.Add(charp2 + "p2");
}
if (flags.useFeaturesCpC5gram || flags.useFeaturesCpC6gram)
{
features.Add(charc3 + "c3");
features.Add(charc2 + charc3 + "c2c3");
}
if (flags.useFeaturesCpC6gram)
{
features.Add(charp3 + "p3");
features.Add(charp3 + charp2 + "p3p2");
}
if (flags.useGoodForNamesCpC)
{
// these 2 features should be distinctively good at biasing from
// picking up a Chinese family name in the p2 or p3 positions:
// familyName X X startWord AND familyName X startWord
// But actually they seem to have negative value.
features.Add(charp2 + "p2");
features.Add(charp3 + "p3");
}
if (flags.useUnicodeType || flags.useUnicodeType4gram || flags.useUnicodeType5gram)
{
features.Add(uTypep + "-" + uTypec + "-" + uTypec2 + "-uType3");
}
if (flags.useUnicodeType4gram || flags.useUnicodeType5gram)
{
features.Add(uTypep2 + "-" + uTypep + "-" + uTypec + "-" + uTypec2 + "-uType4");
}
if (flags.useUnicodeType5gram)
{
features.Add(uTypep2 + "-" + uTypep + "-" + uTypec + "-" + uTypec2 + "-" + uTypec3 + "-uType5");
}
if (flags.useWordUTypeConjunctions2)
{
features.Add(uTypep + charc + "putcc");
features.Add(charp + uTypec + "pccut");
}
if (flags.useWordUTypeConjunctions3)
{
features.Add(uTypep2 + uTypep + charc + "p2utputcc");
features.Add(uTypep + charc + uTypec2 + "putccc2ut");
features.Add(charc + uTypec2 + uTypec3 + "ccc2utc3ut");
}
if (flags.useUnicodeBlock)
{
features.Add(p.GetString <CoreAnnotations.UBlockAnnotation>() + "-" + c.GetString <CoreAnnotations.UBlockAnnotation>() + "-" + c2.GetString <CoreAnnotations.UBlockAnnotation>() + "-uBlock");
}
if (flags.useShapeStrings)
{
if (flags.useShapeStrings1)
{
features.Add(p.GetString <CoreAnnotations.ShapeAnnotation>() + "ps");
features.Add(c.GetString <CoreAnnotations.ShapeAnnotation>() + "cs");
features.Add(c2.GetString <CoreAnnotations.ShapeAnnotation>() + "c2s");
}
if (flags.useShapeStrings3)
{
features.Add(p.GetString <CoreAnnotations.ShapeAnnotation>() + c.GetString <CoreAnnotations.ShapeAnnotation>() + c2.GetString <CoreAnnotations.ShapeAnnotation>() + "pscsc2s");
}
if (flags.useShapeStrings4)
{
features.Add(p2.GetString <CoreAnnotations.ShapeAnnotation>() + p.GetString <CoreAnnotations.ShapeAnnotation>() + c.GetString <CoreAnnotations.ShapeAnnotation>() + c2.GetString <CoreAnnotations.ShapeAnnotation>() + "p2spscsc2s");
}
if (flags.useShapeStrings5)
{
features.Add(p2.GetString <CoreAnnotations.ShapeAnnotation>() + p.GetString <CoreAnnotations.ShapeAnnotation>() + c.GetString <CoreAnnotations.ShapeAnnotation>() + c2.GetString <CoreAnnotations.ShapeAnnotation>() + c3.GetString <CoreAnnotations.ShapeAnnotation
>() + "p2spscsc2sc3s");
}
if (flags.useWordShapeConjunctions2)
{
features.Add(p.GetString <CoreAnnotations.ShapeAnnotation>() + charc + "pscc");
features.Add(charp + c.GetString <CoreAnnotations.ShapeAnnotation>() + "pccs");
}
if (flags.useWordShapeConjunctions3)
{
features.Add(p2.GetString <CoreAnnotations.ShapeAnnotation>() + p.GetString <CoreAnnotations.ShapeAnnotation>() + charc + "p2spscc");
features.Add(p.GetString <CoreAnnotations.ShapeAnnotation>() + charc + c2.GetString <CoreAnnotations.ShapeAnnotation>() + "psccc2s");
features.Add(charc + c2.GetString <CoreAnnotations.ShapeAnnotation>() + c3.GetString <CoreAnnotations.ShapeAnnotation>() + "ccc2sc3s");
}
}
/*
* Radical N-gram features. N is upto 4.
* Smoothing method of N-gram, because there are too many characters in Chinese.
* (It works better than N-gram when they are used individually. less sparse)
*/
char rcharc;
char rcharc2;
char rcharp;
char rcharp2;
if (charc.Length == 0)
{
rcharc = 'n';
}
else
{
rcharc = RadicalMap.GetRadical(charc[0]);
}
if (charc2.Length == 0)
{
rcharc2 = 'n';
}
else
{
rcharc2 = RadicalMap.GetRadical(charc2[0]);
}
if (charp.Length == 0)
{
rcharp = 'n';
}
else
{
rcharp = RadicalMap.GetRadical(charp[0]);
}
if (charp2.Length == 0)
{
rcharp2 = 'n';
}
else
{
rcharp2 = RadicalMap.GetRadical(charp2[0]);
}
if (flags.useRad2)
{
features.Add(rcharc + "rc");
features.Add(rcharc2 + "rc2");
features.Add(rcharp + "rp");
features.Add(rcharp + rcharc + "rprc");
features.Add(rcharc + rcharc2 + "rcrc2");
features.Add(rcharp + rcharc + rcharc2 + "rprcrc2");
}
if (flags.useRad2b)
{
features.Add(rcharc + "rc");
features.Add(rcharc2 + "rc2");
features.Add(rcharp + "rp");
features.Add(rcharp + rcharc + "rprc");
features.Add(rcharc + rcharc2 + "rcrc2");
features.Add(rcharp2 + rcharp + "rp2rp");
}
/* Non-word dictionary: SEEN bi-gram marked as non-word.
* This is frickin' useful. I hadn't realized. CDM Oct 2007.
*/
if (flags.useDict2)
{
NonDict2 nd = new NonDict2(flags);
features.Add(nd.CheckDic(charp + charc, flags) + "nondict");
}
if (flags.useOutDict2)
{
if (outDict == null)
{
CreateOutDict();
}
features.Add(outDict.GetW(charp + charc) + "outdict");
// -1 0
features.Add(outDict.GetW(charc + charc2) + "outdict");
// 0 1
features.Add(outDict.GetW(charp2 + charp) + "outdict");
// -2 -1
features.Add(outDict.GetW(charp2 + charp + charc) + "outdict");
// -2 -1 0
features.Add(outDict.GetW(charp3 + charp2 + charp) + "outdict");
// -3 -2 -1
features.Add(outDict.GetW(charp + charc + charc2) + "outdict");
// -1 0 1
features.Add(outDict.GetW(charc + charc2 + charc3) + "outdict");
// 0 1 2
features.Add(outDict.GetW(charp + charc + charc2 + charc3) + "outdict");
}
// -1 0 1 2
/*
* (CTB/ASBC/HK/PK/MSR) POS information of each characters.
* If a character falls into some function categories,
* it is very likely there is a boundary.
* A lot of Chinese function words belong to single characters.
* This feature is also good for numbers and punctuations.
* DE* are grouped into DE.
*/
if (flags.useCTBChar2 || flags.useASBCChar2 || flags.useHKChar2 || flags.usePKChar2 || flags.useMSRChar2)
{
string[] tagsets;
// the "useChPos" now only works for CTB and PK
if (flags.useChPos)
{
if (flags.useCTBChar2)
{
tagsets = new string[] { "AD", "AS", "BA", "CC", "CD", "CS", "DE", "DT", "ETC", "IJ", "JJ", "LB", "LC", "M", "NN", "NR", "NT", "OD", "P", "PN", "PU", "SB", "SP", "VA", "VC", "VE", "VV" };
}
else
{
if (flags.usePKChar2)
{
//tagsets = new String[]{"r", "j", "t", "a", "nz", "l", "vn", "i", "m", "ns", "nr", "v", "n", "q", "Ng", "b", "d", "nt"};
tagsets = new string[] { "2", "3", "4" };
}
else
{
throw new Exception("only support settings for CTB and PK now.");
}
}
}
else
{
//logger.info("Using Derived features");
tagsets = new string[] { "2", "3", "4" };
}
if (taDetector == null)
{
CreateTADetector();
}
foreach (string tag in tagsets)
{
features.Add(taDetector.CheckDic(tag + "p", charp) + taDetector.CheckDic(tag + "i", charp) + taDetector.CheckDic(tag + "s", charc) + taDetector.CheckInDic(charp) + taDetector.CheckInDic(charc) + tag + "prep-sufc");
}
}
//features.add("|ctbchar2");
/*
* In error analysis, we found English words and numbers are often separated.
* Rule 1: isNumber feature: check if the current and previous char is a number.
* Rule 2: Disambiguation of time point and time duration.
* Rule 3: isEnglish feature: check if the current and previous character is an english letter.
* Rule 4: English name feature: check if the current char is a conjunct pu for English first and last name, since there is no space between two names.
* Most of PUs are a good indicator for word boundary, but - and . is a strong indicator that there is no boundry within a previous , a follow char and it.
*/
if (flags.useRule2)
{
/* Reduplication features */
// previous character == current character
if (charp.Equals(charc))
{
features.Add("11-R2");
}
// previous character == next character
if (charp.Equals(charc2))
{
features.Add("22-R2");
}
// current character == next next character
// fire only when usePk and useHk are both false.
// Notice: this should be (almost) the same as the "22" feature, but we keep it for now.
if (!flags.usePk && !flags.useHk)
{
if (charc.Equals(charc2))
{
features.Add("33-R2");
}
}
char cur1 = ' ';
char cur2 = ' ';
char cur = ' ';
char pre = ' ';
// actually their length must be either 0 or 1
if (charc2.Length > 0)
{
cur1 = charc2[0];
}
if (charc3.Length > 0)
{
cur2 = charc3[0];
}
if (charc.Length > 0)
{
cur = charc[0];
}
if (charp.Length > 0)
{
pre = charp[0];
}
string prer = rcharp.ToString();
// the radical of previous character
Pattern E = Pattern.Compile("[a-zA-Z]");
Pattern N = Pattern.Compile("[0-9]");
Matcher m = E.Matcher(charp);
Matcher ce = E.Matcher(charc);
Matcher pe = E.Matcher(charp2);
Matcher cn = N.Matcher(charc);
Matcher pn = N.Matcher(charp2);
// if current and previous characters are numbers...
if (cur >= '0' && cur <= '9' && pre >= '0' && pre <= '9')
{
if (cur == '9' && pre == '1' && cur1 == '9' && cur2 >= '0' && cur2 <= '9')
{
//199x
features.Add("YR-R2");
}
else
{
features.Add("2N-R2");
}
}
else
{
// if current and previous characters are not both numbers
// but previous char is a number
// i.e. patterns like "1N" , "2A", etc
if (pre >= '0' && pre <= '9')
{
features.Add("1N-R2");
}
else
{
// if previous character is an English character
if (m.Matches())
{
features.Add("E-R2");
}
else
{
// if the previous character contains no radical (and it exist)
if (prer.Equals(".") && charp.Length == 1)
{
if (ce.Matches())
{
features.Add("PU+E-R2");
}
if (pe.Matches())
{
features.Add("E+PU-R2");
}
if (cn.Matches())
{
features.Add("PU+N-R2");
}
if (pn.Matches())
{
features.Add("N+PU-R2");
}
features.Add("PU-R2");
}
}
}
}
string engType = IsEnglish(charp, charc);
string engPU = IsEngPU(charp);
if (!engType.Equals(string.Empty))
{
features.Add(engType);
}
if (!engPU.Equals(string.Empty) && !engType.Equals(string.Empty))
{
StringBuilder sb = new StringBuilder();
sb.Append(engPU).Append(engType).Append("R2");
features.Add(sb.ToString());
}
}
//end of use rule
// features using "Character.getType" information!
string origS = c.GetString <CoreAnnotations.OriginalCharAnnotation>();
char origC = ' ';
if (origS.Length > 0)
{
origC = origS[0];
}
int type = char.GetType(origC);
switch (type)
{
case char.UppercaseLetter:
case char.LowercaseLetter:
{
// A-Z and full-width A-Z
// a-z and full-width a-z
features.Add("CHARTYPE-LETTER");
break;
}
case char.DecimalDigitNumber:
{
features.Add("CHARTYPE-DECIMAL_DIGIT_NUMBER");
break;
}
case char.OtherLetter:
{
// mostly chinese chars
features.Add("CHARTYPE-OTHER_LETTER");
break;
}
default:
{
// other types
features.Add("CHARTYPE-MISC");
break;
}
}
features.Add("cliqueCpC");
return(features);
}
protected internal virtual ICollection <string> FeaturesC <_T0>(PaddedList <_T0> cInfo, int loc)
where _T0 : CoreLabel
{
ICollection <string> features = new List <string>();
CoreLabel c = cInfo[loc];
CoreLabel c2 = cInfo[loc + 1];
CoreLabel c3 = cInfo[loc + 2];
CoreLabel p = cInfo[loc - 1];
CoreLabel p2 = cInfo[loc - 2];
CoreLabel p3 = cInfo[loc - 3];
string charc = c.GetString <CoreAnnotations.CharAnnotation>();
string charc2 = c2.GetString <CoreAnnotations.CharAnnotation>();
string charc3 = c3.GetString <CoreAnnotations.CharAnnotation>();
string charp = p.GetString <CoreAnnotations.CharAnnotation>();
string charp2 = p2.GetString <CoreAnnotations.CharAnnotation>();
string charp3 = p3.GetString <CoreAnnotations.CharAnnotation>();
int cI = c.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypec = (cI != null ? cI.ToString() : string.Empty);
int c2I = c2.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypec2 = (c2I != null ? c2I.ToString() : string.Empty);
int c3I = c3.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypec3 = (c3I != null ? c3I.ToString() : string.Empty);
int pI = p.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypep = (pI != null ? pI.ToString() : string.Empty);
int p2I = p2.Get(typeof(CoreAnnotations.UTypeAnnotation));
string uTypep2 = (p2I != null ? p2I.ToString() : string.Empty);
/* N-gram features. N is upto 2. */
if (flags.useWord1)
{
// features.add(charc +"c");
// features.add(charc2+"c2");
// features.add(charp +"p");
// features.add(charp + charc +"pc");
// features.add(charc + charc2 +"cc2");
// cdm: need hyphen so you can see which of charp or charc2 is null....
// features.add(charp + "-" + charc2 + "pc2");
features.Add(charc + "::c");
features.Add(charc2 + "::c2");
features.Add(charp + "::p");
features.Add(charp2 + "::p2");
// trying to restore the features that Huishin described in SIGHAN 2005 paper
features.Add(charc + charc2 + "::cn");
features.Add(charc + charc3 + "::cn2");
features.Add(charp + charc + "::pc");
features.Add(charp + charc2 + "::pn");
features.Add(charp2 + charp + "::p2p");
features.Add(charp2 + charc + "::p2c");
features.Add(charc2 + charc + "::n2c");
}
if (flags.dictionary != null || flags.serializedDictionary != null)
{
DictionaryFeaturesC(typeof(CoreAnnotations.LBeginAnnotation), typeof(CoreAnnotations.LMiddleAnnotation), typeof(CoreAnnotations.LEndAnnotation), string.Empty, features, p, c, c2);
}
if (flags.dictionary2 != null)
{
DictionaryFeaturesC(typeof(CoreAnnotations.D2_LBeginAnnotation), typeof(CoreAnnotations.D2_LMiddleAnnotation), typeof(CoreAnnotations.D2_LEndAnnotation), "-D2-", features, p, c, c2);
}
if (flags.useFeaturesC4gram || flags.useFeaturesC5gram || flags.useFeaturesC6gram)
{
features.Add(charp2 + charp + "p2p");
features.Add(charp2 + "p2");
}
if (flags.useFeaturesC5gram || flags.useFeaturesC6gram)
{
features.Add(charc3 + "c3");
features.Add(charc2 + charc3 + "c2c3");
}
if (flags.useFeaturesC6gram)
{
features.Add(charp3 + "p3");
features.Add(charp3 + charp2 + "p3p2");
}
if (flags.useUnicodeType || flags.useUnicodeType4gram || flags.useUnicodeType5gram)
{
features.Add(uTypep + "-" + uTypec + "-" + uTypec2 + "-uType3");
}
if (flags.useUnicodeType4gram || flags.useUnicodeType5gram)
{
features.Add(uTypep2 + "-" + uTypep + "-" + uTypec + "-" + uTypec2 + "-uType4");
}
if (flags.useUnicodeType5gram)
{
features.Add(uTypep2 + "-" + uTypep + "-" + uTypec + "-" + uTypec2 + "-" + uTypec3 + "-uType5");
}
if (flags.useUnicodeBlock)
{
features.Add(p.GetString <CoreAnnotations.UBlockAnnotation>() + "-" + c.GetString <CoreAnnotations.UBlockAnnotation>() + "-" + c2.GetString <CoreAnnotations.UBlockAnnotation>() + "-uBlock");
}
if (flags.useShapeStrings)
{
if (flags.useShapeStrings1)
{
features.Add(p.GetString <CoreAnnotations.ShapeAnnotation>() + "ps");
features.Add(c.GetString <CoreAnnotations.ShapeAnnotation>() + "cs");
features.Add(c2.GetString <CoreAnnotations.ShapeAnnotation>() + "c2s");
}
if (flags.useShapeStrings3)
{
features.Add(p.GetString <CoreAnnotations.ShapeAnnotation>() + c.GetString <CoreAnnotations.ShapeAnnotation>() + c2.GetString <CoreAnnotations.ShapeAnnotation>() + "pscsc2s");
}
if (flags.useShapeStrings4)
{
features.Add(p2.GetString <CoreAnnotations.ShapeAnnotation>() + p.GetString <CoreAnnotations.ShapeAnnotation>() + c.GetString <CoreAnnotations.ShapeAnnotation>() + c2.GetString <CoreAnnotations.ShapeAnnotation>() + "p2spscsc2s");
}
if (flags.useShapeStrings5)
{
features.Add(p2.GetString <CoreAnnotations.ShapeAnnotation>() + p.GetString <CoreAnnotations.ShapeAnnotation>() + c.GetString <CoreAnnotations.ShapeAnnotation>() + c2.GetString <CoreAnnotations.ShapeAnnotation>() + c3.GetString <CoreAnnotations.ShapeAnnotation
>() + "p2spscsc2sc3s");
}
}
features.Add("cliqueC");
return(features);
}
// All the tags we need
// Patterns we need
// In theory 块 钱 should be separated by segmenter, but just in case segmenter fails
// TODO(yuhao): Need to add support for 块 钱, 毛 钱, 角 钱, 角, 五 块 二
// This only works when POS = NT
// This is used to capture a special case of date in Chinese: 70 后 or 七零 后
// order it by number of characters DESC for handy one-by-one matching of string suffix
/// <summary>Use a set of heuristic rules to assign NER tags to tokens.</summary>
/// <param name="document">
/// A
/// <see cref="System.Collections.IList{E}"/>
/// of something that extends
/// <see cref="Edu.Stanford.Nlp.Util.ICoreMap"/>
/// .
/// </param>
/// <returns/>
public override IList<CoreLabel> Classify(IList<CoreLabel> document)
{
// The actual implementation of the classifier
PaddedList<CoreLabel> pl = new PaddedList<CoreLabel>(document, pad);
for (int i = 0; i < sz; i++)
{
CoreLabel me = pl[i];
CoreLabel prev = pl[i - 1];
CoreLabel next = pl[i + 1];
// by default set to be "O"
me.Set(typeof(CoreAnnotations.AnswerAnnotation), flags.backgroundSymbol);
// If current word is OD, label it as ORDINAL
if (me.GetString<CoreAnnotations.PartOfSpeechAnnotation>().Equals("OD"))
{
me.Set(typeof(CoreAnnotations.AnswerAnnotation), OrdinalTag);
}
else
{
if (CurrencyWordPattern.Matcher(me.Word()).Matches() && prev.GetString<CoreAnnotations.PartOfSpeechAnnotation>().Equals("CD"))
{
// If current word is currency word and prev word is a CD
me.Set(typeof(CoreAnnotations.AnswerAnnotation), MoneyTag);
}
else
{
if (me.GetString<CoreAnnotations.PartOfSpeechAnnotation>().Equals("CD"))
{
// TODO(yuhao): Need to support Chinese captial numbers like 叁拾 (This won't be POS-tagged as CD).
// If current word is a CD
if (PercentWordPattern1.Matcher(me.Word()).Matches() || PercentWordPattern2.Matcher(me.Word()).Matches())
{
// If current word is a percent
me.Set(typeof(CoreAnnotations.AnswerAnnotation), PercentTag);
}
else
{
if (RightScanFindsMoneyWord(pl, i))
{
// If one the right finds a currency word
me.Set(typeof(CoreAnnotations.AnswerAnnotation), MoneyTag);
}
else
{
if (me.Word().Length == 2 && ChineseAndArabicNumeralsPattern.Matcher(me.Word()).Matches() && DateAgeLocalizer.Equals(next.Word()))
{
// This is to extract a special case of DATE: 70 后 or 七零 后
me.Set(typeof(CoreAnnotations.AnswerAnnotation), DateTag);
}
else
{
// Otherwise we should safely label it as NUMBER
me.Set(typeof(CoreAnnotations.AnswerAnnotation), NumberTag);
}
}
}
}
else
{
if (me.GetString<CoreAnnotations.PartOfSpeechAnnotation>().Equals("NT"))
{
// If current word is a NT (temporal noun)
if (DatePattern1.Matcher(me.Word()).Matches() || DatePattern2.Matcher(me.Word()).Matches() || DatePattern3.Matcher(me.Word()).Matches() || DatePattern4.Matcher(me.Word()).Matches() || DatePattern5.Matcher(me.Word()).Matches() || DateWords.Contains
(me.Word()))
{
me.Set(typeof(CoreAnnotations.AnswerAnnotation), DateTag);
}
else
{
if (TimePattern1.Matcher(me.Word()).Matches() || TimeWords.Contains(me.Word()))
{
me.Set(typeof(CoreAnnotations.AnswerAnnotation), TimeTag);
}
else
{
// TIME may have more variants (really?) so always add as TIME by default
me.Set(typeof(CoreAnnotations.AnswerAnnotation), TimeTag);
}
}
}
else
{
if (DateAgeLocalizer.Equals(me.Word()) && prev.Word() != null && prev.Word().Length == 2 && ChineseAndArabicNumeralsPattern.Matcher(prev.Word()).Matches())
{
// Label 后 as DATE if the sequence is 70 后 or 七零 后
me.Set(typeof(CoreAnnotations.AnswerAnnotation), DateTag);
}
}
}
}
}
}
return document;
}
public override ICollection <string> GetCliqueFeatures(PaddedList <In> info, int position, Clique clique)
{
IList <string> features = new List <string>(Arrays.AsList(info[position].Word().Split(" ")));
return(features);
}
/// <summary>
/// This method returns a
/// <see cref="System.Collections.ICollection{E}"/>
/// of the features
/// calculated for the word at the specified position in info (the list of
/// words) for the specified
/// <see cref="Clique"/>
/// .
/// It should return the actual String features, <b>NOT</b> wrapped in any
/// other object, as the wrapping
/// will be done automatically.
/// Because it takes a
/// <see cref="Edu.Stanford.Nlp.Util.PaddedList{E}"/>
/// you don't
/// need to worry about indices which are outside of the list.
/// </summary>
/// <param name="info">A PaddedList of the feature-value pairs</param>
/// <param name="position">The current position to extract features at</param>
/// <param name="clique">
/// The particular clique for which to extract features. It
/// should be a member of the knownCliques list.
/// </param>
/// <returns>
/// A
/// <see cref="System.Collections.ICollection{E}"/>
/// of the features
/// calculated for the word at the specified position in info.
/// </returns>
public abstract ICollection <string> GetCliqueFeatures(PaddedList <In> info, int position, Clique clique);
public virtual ICollection <string> FeaturesCpC(PaddedList <IN> cInfo, int loc)
{
ICollection <string> features = new List <string>();
CoreLabel c = cInfo[loc];
CoreLabel c1 = cInfo[loc + 1];
CoreLabel c2 = cInfo[loc + 2];
CoreLabel c3 = cInfo[loc + 3];
CoreLabel p = cInfo[loc - 1];
CoreLabel p2 = cInfo[loc - 2];
CoreLabel p3 = cInfo[loc - 3];
string charc = c.Get(typeof(CoreAnnotations.CharAnnotation));
if (charc == null)
{
charc = string.Empty;
}
string charc1 = c1.Get(typeof(CoreAnnotations.CharAnnotation));
if (charc1 == null)
{
charc1 = string.Empty;
}
string charc2 = c2.Get(typeof(CoreAnnotations.CharAnnotation));
if (charc2 == null)
{
charc2 = string.Empty;
}
string charc3 = c3.Get(typeof(CoreAnnotations.CharAnnotation));
if (charc3 == null)
{
charc3 = string.Empty;
}
string charp = p.Get(typeof(CoreAnnotations.CharAnnotation));
if (charp == null)
{
charp = string.Empty;
}
string charp2 = p2.Get(typeof(CoreAnnotations.CharAnnotation));
if (charp2 == null)
{
charp2 = string.Empty;
}
string charp3 = p3.Get(typeof(CoreAnnotations.CharAnnotation));
if (charp3 == null)
{
charp3 = string.Empty;
}
/*
* N-gram features. N is upto 2.
*/
if (flags.useWord2)
{
// features.add(charc +"c");
// features.add(charc1+"c1");
// features.add(charp +"p");
// features.add(charp +charc +"pc");
// if( flags.useMsr ){
// features.add(charc +charc1 +"cc1");
// features.add(charp + charc1 +"pc1");
// }
features.Add(charc + "::c");
features.Add(charc1 + "::c1");
features.Add(charp + "::p");
features.Add(charp2 + "::p2");
// trying to restore the features that Huishin described in SIGHAN 2005 paper
features.Add(charc + charc1 + "::cn");
features.Add(charp + charc + "::pc");
features.Add(charp + charc1 + "::pn");
features.Add(charp2 + charp + "::p2p");
features.Add(charp2 + charc + "::p2c");
features.Add(charc2 + charc + "::n2c");
features.Add("|word2");
}
/*
* Radical N-gram features. N is upto 4.
* Smoothing method of N-gram, because there are too many characters in Chinese.
* (It works better than N-gram when they are used individually. less sparse)
*/
char rcharc;
char rcharc1;
char rcharc2;
char rcharc3;
char rcharp;
char rcharp1;
char rcharp2;
char rcharp3;
if (charc.Length == 0)
{
rcharc = 'n';
}
else
{
rcharc = RadicalMap.GetRadical(charc[0]);
}
if (charc1.Length == 0)
{
rcharc1 = 'n';
}
else
{
rcharc1 = RadicalMap.GetRadical(charc1[0]);
}
if (charc2.Length == 0)
{
rcharc2 = 'n';
}
else
{
rcharc2 = RadicalMap.GetRadical(charc2[0]);
}
if (charc3.Length == 0)
{
rcharc3 = 'n';
}
else
{
rcharc3 = RadicalMap.GetRadical(charc3[0]);
}
if (charp.Length == 0)
{
rcharp = 'n';
}
else
{
rcharp = RadicalMap.GetRadical(charp[0]);
}
if (charp2.Length == 0)
{
rcharp2 = 'n';
}
else
{
rcharp2 = RadicalMap.GetRadical(charp2[0]);
}
if (charp3.Length == 0)
{
rcharp3 = 'n';
}
else
{
rcharp3 = RadicalMap.GetRadical(charp3[0]);
}
if (flags.useRad2)
{
features.Add(rcharc + "rc");
features.Add(rcharc1 + "rc1");
features.Add(rcharp + "rp");
features.Add(rcharp + rcharc + "rpc");
features.Add(rcharc + rcharc1 + "rcc1");
features.Add(rcharp + rcharc + rcharc1 + "rpcc1");
features.Add("|rad2");
}
/* non-word dictionary:SEEM bi-gram marked as non-word */
if (flags.useDict2)
{
NonDict2 nd = new NonDict2(flags);
features.Add(nd.CheckDic(charp + charc, flags) + "nondict");
features.Add("|useDict2");
}
if (flags.useOutDict2)
{
if (outDict == null)
{
logger.Info("reading " + flags.outDict2 + " as a seen lexicon");
outDict = new CorpusDictionary(flags.outDict2, true);
}
features.Add(outDict.GetW(charp + charc) + "outdict");
// -1 0
features.Add(outDict.GetW(charc + charc1) + "outdict");
// 0 1
features.Add(outDict.GetW(charp2 + charp) + "outdict");
// -2 -1
features.Add(outDict.GetW(charp2 + charp + charc) + "outdict");
// -2 -1 0
features.Add(outDict.GetW(charp3 + charp2 + charp) + "outdict");
// -3 -2 -1
features.Add(outDict.GetW(charp + charc + charc1) + "outdict");
// -1 0 1
features.Add(outDict.GetW(charc + charc1 + charc2) + "outdict");
// 0 1 2
features.Add(outDict.GetW(charp + charc + charc1 + charc2) + "outdict");
}
// -1 0 1 2
/*
* (CTB/ASBC/HK/PK/MSR) POS information of each characters.
* If a character falls into some function categories,
* it is very likely there is a boundary.
* A lot of Chinese function words belong to single characters.
* This feature is also good for numbers and punctuations.
* DE* are grouped into DE.
*/
if (flags.useCTBChar2 || flags.useASBCChar2 || flags.useHKChar2 || flags.usePKChar2 || flags.useMSRChar2)
{
string[] tagsets;
// the "useChPos" now only works for CTB and PK
if (flags.useChPos)
{
if (flags.useCTBChar2)
{
tagsets = new string[] { "AD", "AS", "BA", "CC", "CD", "CS", "DE", "DT", "ETC", "IJ", "JJ", "LB", "LC", "M", "NN", "NR", "NT", "OD", "P", "PN", "PU", "SB", "SP", "VA", "VC", "VE", "VV" };
}
else
{
if (flags.usePKChar2)
{
//tagsets = new String[]{"r", "j", "t", "a", "nz", "l", "vn", "i", "m", "ns", "nr", "v", "n", "q", "Ng", "b", "d", "nt"};
tagsets = new string[] { "2", "3", "4" };
}
else
{
throw new Exception("only support settings for CTB and PK now.");
}
}
}
else
{
//logger.info("Using Derived features");
tagsets = new string[] { "2", "3", "4" };
}
if (taDetector == null)
{
taDetector = new TagAffixDetector(flags);
}
foreach (string tagset in tagsets)
{
features.Add(taDetector.CheckDic(tagset + "p", charp) + taDetector.CheckDic(tagset + "i", charp) + taDetector.CheckDic(tagset + "s", charc) + taDetector.CheckInDic(charp) + taDetector.CheckInDic(charc) + tagset + "prep-sufc");
}
}
// features.add("|ctbchar2"); // Added a constant feature several times!!
/*
* In error analysis, we found English words and numbers are often separated.
* Rule 1: isNumber feature: check if the current and previous char is a number.
* Rule 2: Disambiguation of time point and time duration.
* Rule 3: isEnglish feature: check if the current and previous character is an english letter.
* Rule 4: English name feature: check if the current char is a conjunct pu for English first and last name, since there is no space between two names.
* Most of PUs are a good indicator for word boundary, but - and . is a strong indicator that there is no boundry within a previous , a follow char and it.
*/
if (flags.useRule2)
{
/* Reduplication features */
// previous character == current character
if (charp.Equals(charc))
{
features.Add("11");
}
// previous character == next character
if (charp.Equals(charc1))
{
features.Add("22");
}
// current character == next next character
// fire only when usePk and useHk are both false.
// Notice: this should be (almost) the same as the "22" feature, but we keep it for now.
if (!flags.usePk && !flags.useHk)
{
if (charc.Equals(charc2))
{
features.Add("33");
}
}
char cur1 = ' ';
char cur2 = ' ';
char cur = ' ';
char pre = ' ';
// actually their length must be either 0 or 1
if (charc1.Length > 0)
{
cur1 = charc1[0];
}
if (charc2.Length > 0)
{
cur2 = charc2[0];
}
if (charc.Length > 0)
{
cur = charc[0];
}
if (charp.Length > 0)
{
pre = charp[0];
}
string prer = rcharp.ToString();
// the radical of previous character
Pattern E = Pattern.Compile("[a-zA-Z]");
Pattern N = Pattern.Compile("[0-9]");
Matcher m = E.Matcher(charp);
Matcher ce = E.Matcher(charc);
Matcher pe = E.Matcher(charp2);
Matcher cn = N.Matcher(charc);
Matcher pn = N.Matcher(charp2);
// if current and previous characters are numbers...
if (cur >= '0' && cur <= '9' && pre >= '0' && pre <= '9')
{
if (cur == '9' && pre == '1' && cur1 == '9' && cur2 >= '0' && cur2 <= '9')
{
//199x
features.Add("YR");
}
else
{
features.Add("2N");
}
}
else
{
// if current and previous characters are not both numbers
// but previous char is a number
// i.e. patterns like "1N" , "2A", etc
if (pre >= '0' && pre <= '9')
{
features.Add("1N");
}
else
{
// if previous character is an English character
if (m.Matches())
{
features.Add("E");
}
else
{
// if the previous character contains no radical (and it exist)
if (prer.Equals(".") && charp.Length == 1)
{
// fire only when usePk and useHk are both false. Not sure why. -pichuan
if (!flags.useHk && !flags.usePk)
{
if (ce.Matches())
{
features.Add("PU+E");
}
if (pe.Matches())
{
features.Add("E+PU");
}
if (cn.Matches())
{
features.Add("PU+N");
}
if (pn.Matches())
{
features.Add("N+PU");
}
}
features.Add("PU");
}
}
}
}
string engType = IsEnglish(charp, charc);
string engPU = IsEngPU(charp);
if (!engType.Equals(string.Empty))
{
features.Add(engType);
}
if (!engPU.Equals(string.Empty) && !engType.Equals(string.Empty))
{
features.Add(engPU + engType);
}
}
//end of use rule
// features using "Character.getType" information!
string origS = c.Get(typeof(CoreAnnotations.OriginalCharAnnotation));
char origC = ' ';
if (origS.Length > 0)
{
origC = origS[0];
}
int type = char.GetType(origC);
switch (type)
{
case char.UppercaseLetter:
case char.LowercaseLetter:
{
// A-Z and full-width A-Z
// a-z and full-width a-z
features.Add("CHARTYPE-LETTER");
break;
}
case char.DecimalDigitNumber:
{
features.Add("CHARTYPE-DECIMAL_DIGIT_NUMBER");
break;
}
case char.OtherLetter:
{
// mostly chinese chars
features.Add("CHARTYPE-OTHER_LETTER");
break;
}
default:
{
// other types
features.Add("CHARTYPE-MISC");
break;
}
}
return(features);
}
