private ChineseCharacterBasedLexicon.Symbol UnknownCharClass(ChineseCharacterBasedLexicon.Symbol ch)
{
if (useUnknownCharacterModel)
{
return(new ChineseCharacterBasedLexicon.Symbol(char.ToString(RadicalMap.GetRadical(ch.GetCh()))).Intern());
}
else
{
return(ChineseCharacterBasedLexicon.Symbol.Unknown);
}
}
public static void PrintStats(ICollection <Tree> trees, PrintWriter pw)
{
ClassicCounter <int> wordLengthCounter = new ClassicCounter <int>();
ClassicCounter <TaggedWord> wordCounter = new ClassicCounter <TaggedWord>();
ClassicCounter <ChineseCharacterBasedLexicon.Symbol> charCounter = new ClassicCounter <ChineseCharacterBasedLexicon.Symbol>();
int counter = 0;
foreach (Tree tree in trees)
{
counter++;
IList <TaggedWord> taggedWords = tree.TaggedYield();
foreach (TaggedWord taggedWord in taggedWords)
{
string word = taggedWord.Word();
if (word.Equals(LexiconConstants.Boundary))
{
continue;
}
wordCounter.IncrementCount(taggedWord);
wordLengthCounter.IncrementCount(int.Parse(word.Length));
for (int j = 0; j < length; j++)
{
ChineseCharacterBasedLexicon.Symbol sym = ChineseCharacterBasedLexicon.Symbol.CannonicalSymbol(word[j]);
charCounter.IncrementCount(sym);
}
charCounter.IncrementCount(ChineseCharacterBasedLexicon.Symbol.EndWord);
}
}
ICollection <ChineseCharacterBasedLexicon.Symbol> singletonChars = Counters.KeysBelow(charCounter, 1.5);
ICollection <TaggedWord> singletonWords = Counters.KeysBelow(wordCounter, 1.5);
ClassicCounter <string> singletonWordPOSes = new ClassicCounter <string>();
foreach (TaggedWord taggedWord_1 in singletonWords)
{
singletonWordPOSes.IncrementCount(taggedWord_1.Tag());
}
Distribution <string> singletonWordPOSDist = Distribution.GetDistribution(singletonWordPOSes);
ClassicCounter <char> singletonCharRads = new ClassicCounter <char>();
foreach (ChineseCharacterBasedLexicon.Symbol s in singletonChars)
{
singletonCharRads.IncrementCount(char.ValueOf(RadicalMap.GetRadical(s.GetCh())));
}
Distribution <char> singletonCharRadDist = Distribution.GetDistribution(singletonCharRads);
Distribution <int> wordLengthDist = Distribution.GetDistribution(wordLengthCounter);
NumberFormat percent = new DecimalFormat("##.##%");
pw.Println("There are " + singletonChars.Count + " singleton chars out of " + (int)charCounter.TotalCount() + " tokens and " + charCounter.Size() + " types found in " + counter + " trees.");
pw.Println("Thus singletonChars comprise " + percent.Format(singletonChars.Count / charCounter.TotalCount()) + " of tokens and " + percent.Format((double)singletonChars.Count / charCounter.Size()) + " of types.");
pw.Println();
pw.Println("There are " + singletonWords.Count + " singleton words out of " + (int)wordCounter.TotalCount() + " tokens and " + wordCounter.Size() + " types.");
pw.Println("Thus singletonWords comprise " + percent.Format(singletonWords.Count / wordCounter.TotalCount()) + " of tokens and " + percent.Format((double)singletonWords.Count / wordCounter.Size()) + " of types.");
pw.Println();
pw.Println("Distribution over singleton word POS:");
pw.Println(singletonWordPOSDist.ToString());
pw.Println();
pw.Println("Distribution over singleton char radicals:");
pw.Println(singletonCharRadDist.ToString());
pw.Println();
pw.Println("Distribution over word length:");
pw.Println(wordLengthDist);
}
//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);
}
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);
}
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);
}
public virtual ICollection <string> MakeFeatures(string word)
{
IList <string> features = new List <string>();
if (morpho)
{
foreach (KeyValuePair <string, ICollection <char> > e in cmfs.GetSingletonFeatures())
{
if (e.Value.Contains(word[0]))
{
features.Add(e.Key + "-1");
}
}
// Hooray for generics!!! :-)
foreach (KeyValuePair <string, Pair <ICollection <char>, ICollection <char> > > e_1 in cmfs.GetAffixFeatures())
{
bool both = false;
if (e_1.Value.First().Contains(word[0]))
{
features.Add(e_1.Key + "-P");
both = true;
}
if (e_1.Value.Second().Contains(word[word.Length - 1]))
{
features.Add(e_1.Key + "-S");
}
else
{
both = false;
}
if (both && mildConjunctions && !conjunctions)
{
features.Add(e_1.Key + "-PS");
}
}
if (conjunctions)
{
int max = features.Count;
for (int i = 1; i < max; i++)
{
string s1 = features[i];
for (int j = 0; j < i; j++)
{
string s2 = features[j];
features.Add(s1 + "&&" + s2);
}
}
}
}
if (!turnOffWordFeatures)
{
features.Add(word + "-W");
}
if (rads)
{
features.Add(RadicalMap.GetRadical(word[0]) + "-FR");
features.Add(RadicalMap.GetRadical(word[word.Length - 1]) + "-LR");
for (int i = 0; i < word.Length; i++)
{
features.Add(RadicalMap.GetRadical(word[i]) + "-CR");
}
}
if (chars)
{
// first and last chars
features.Add(word[0] + "-FC");
features.Add(word[word.Length - 1] + "-LC");
for (int i = 0; i < word.Length; i++)
{
features.Add(word[i] + "-CC");
}
if (bigrams && word.Length > 1)
{
features.Add(Sharpen.Runtime.Substring(word, 0, 2) + "-FB");
features.Add(Sharpen.Runtime.Substring(word, word.Length - 2) + "-LB");
for (int i_1 = 2; i_1 <= word.Length; i_1++)
{
features.Add(Sharpen.Runtime.Substring(word, i_1 - 2, i_1) + "-CB");
}
}
}
if (useLength)
{
int lengthBin = word.Length;
if (lengthBin >= 5)
{
if (lengthBin >= 8)
{
lengthBin = 8;
}
else
{
lengthBin = 5;
}
}
features.Add(word.Length + "-L");
}
if (useFreq && !turnOffWordFeatures)
{
int freq = wordCounter.GetIntCount(word);
int freqBin;
if (freq <= 1)
{
freqBin = 0;
}
else
{
if (freq <= 3)
{
freqBin = 1;
}
else
{
if (freq <= 6)
{
freqBin = 2;
}
else
{
if (freq <= 15)
{
freqBin = 3;
}
else
{
if (freq <= 50)
{
freqBin = 4;
}
else
{
freqBin = 5;
}
}
}
}
}
features.Add(freqBin + "-FQ");
}
features.Add("PR");
if (threshedFeatures != null)
{
for (IEnumerator <string> iter = features.GetEnumerator(); iter.MoveNext();)
{
string s = iter.Current;
if (!threshedFeatures.Contains(s))
{
iter.Remove();
}
}
}
return(features);
}