/**
* 1) Performs re-ranking using a linear function, which takes into account only three features:
* newRank = 1 * oldRank + 10 * underconstrained_query + 10 * incosistent_tense; <br>
* 2) Takes the hypothesis which is smallest in (new) rank as the best one. <br>
* 3) Computes the semantic error rate (SER) as follows: -- a recognition is considered as
* semantically correct if the dialogue move representation it produces is both a) non-null and b)
* the same as the one that would have been produced from a perfect recognition result. -- if
* dialogue move is not the same: then counts the number of deletions/insertions required in order
* to obtain the perfect dialogue move.
*
* @deprecated Use instead the linear re-ranking with more than 3 feat and SER manually defined
* (available in xml file)
* @param xmlFileName - input xml file containing both the n-best hypothesis and the reference
* transcription.
* @return
* @throws Exception
*/
public float[] getER4LinearReranking3Feat(String xmlFileName) throws Exception {
// read the xml file in order to get the utterance transcripts
try {
Document d =
new org.jdom.input.SAXBuilder().build(new File(xmlFileName)); // PARSE THE XML FILE
java.util.List nbestList = d.getRootElement().getChildren("nbest_data");
float[] serArray = new float[nbestList.size()];
int noUtt = 0;
int minNewRankID = 1;
for (int i = 0; i < nbestList.size(); i++) {
Element nbestElem = (Element) nbestList.get(i);
noUtt++;
// In order to COMPUTE SEMANTIC ERROR RATE (ser),
// get the dialogue_move feature value for the correct transcription
// dialogue_move
Element dmElem = nbestElem.getChild("dialogue_move");
String refDM = "";
if (dmElem != null) if (!dmElem.getValue().equalsIgnoreCase("")) refDM = dmElem.getValue();
// In the xml tree: find hyp_transcription,
// i.e. the transcription corresponding to the 1-rank predicted hypothesis
java.util.List recList = nbestElem.getChildren("recognition");
// PERFORM LINEAR RE-RANKING
int minNewRank = 100;
for (int j = 1; j < recList.size(); j++) {
Element recElem = (Element) recList.get(j);
int rank = new Integer(recElem.getChild("rank").getValue()).intValue();
int uq = new Integer(recElem.getChild("underconstrained_query").getValue()).intValue();
int it = new Integer(recElem.getChild("inconsistent_tense").getValue()).intValue();
int newRank = rank + 10 * uq + 10 * it;
if (newRank < minNewRank) {
minNewRank = newRank;
minNewRankID = j;
}
}
Element recElem = (Element) recList.get(minNewRankID);
Element dm4recElem = recElem.getChild("dialogue_move");
String dm4rec = "";
if (dm4recElem != null)
if (!dm4recElem.getValue().equalsIgnoreCase("")) dm4rec = dm4recElem.getValue();
WordErrorRate wer = new WordErrorRate(refDM, dm4rec, this.wordDeliminator);
serArray[i] = wer.computeNumerator();
}
return serArray;
} catch (IOException eIO) {
eIO.printStackTrace();
} catch (JDOMException eJDOM) {
eJDOM.printStackTrace();
}
return null;
}
/**
* Compute the best semantic error rate that can be achieved starting from n-best hypothesis.
*
* @param xmlFileName
* @return a list of error rates corresponding to each utterance
* @throws Exception
*/
public float[] getBestERCanBeAchieved(String xmlFileName) throws Exception {
// read the xml file in order to get the utterance transcripts
try {
Document d =
new org.jdom.input.SAXBuilder().build(new File(xmlFileName)); // PARSE THE XML FILE
java.util.List nbestList = d.getRootElement().getChildren("nbest_data");
float[] serArray = new float[nbestList.size()];
int noUtt = 0;
for (int i = 0; i < nbestList.size(); i++) {
Element nbestElem = (Element) nbestList.get(i);
noUtt++;
// In order to COMPUTE SEMANTIC ERROR RATE (SER)
java.util.List recList = nbestElem.getChildren("recognition");
float bestSER = 10000;
// (start with k=1 since we skip first value in recList which corresponds to the correct
// transcription)
for (int k = 1; k < recList.size(); k++) {
Element recElem = (Element) recList.get(k);
Element semCorrectElem = recElem.getChild("semantically_correct");
String semCorrect = semCorrectElem.getValue();
int ser;
if (semCorrect.equalsIgnoreCase("good")) ser = 0;
else ser = 1;
if (bestSER > ser) {
bestSER = ser;
}
} // end for k
serArray[i] = bestSER;
}
return serArray;
} catch (IOException eIO) {
eIO.printStackTrace();
} catch (JDOMException eJDOM) {
eJDOM.printStackTrace();
}
return null;
}
/**
* Gets the SER for a naive algorithm which does random re-ranking.
*
* @param xmlFileName
* @param svmPredFileName
* @return
* @throws Exception
*/
public float[] getER4RandomHyp(String xmlFileName) throws Exception {
// read the xml file in order to get the utterance transcripts
try {
Document d =
new org.jdom.input.SAXBuilder().build(new File(xmlFileName)); // PARSE THE XML FILE
java.util.List nbestList = d.getRootElement().getChildren("nbest_data");
float[] serArray = new float[nbestList.size()];
int noUtt = 0;
java.util.Random rand = new java.util.Random();
for (int i = 0; i < nbestList.size(); i++) {
Element nbestElem = (Element) nbestList.get(i);
noUtt++;
java.util.List recList = nbestElem.getChildren("recognition");
int bestRankIdx = rand.nextInt(recList.size() - 1) + 1;
// System.out.println("best rank = " + bestRankIdx);
if (recList.size() < bestRankIdx)
System.out.println("Less than " + bestRankIdx + " recognitions !!! ");
Element recElem = (Element) recList.get(bestRankIdx);
Element semCorrectElem = recElem.getChild("semantically_correct");
String semCorrect = semCorrectElem.getValue();
if (semCorrect.equalsIgnoreCase("good")) serArray[i] = 0;
else serArray[i] = 1;
}
return serArray;
} catch (IOException eIO) {
eIO.printStackTrace();
} catch (JDOMException eJDOM) {
eJDOM.printStackTrace();
} catch (NullPointerException nullE) {
nullE.printStackTrace();
}
return null;
}
// ------------------------------------------------------------------------------------------------
// 描述:
// 设计: Skyline(2001.12.29)
// 实现: Skyline
// 修改:
// ------------------------------------------------------------------------------------------------
public void InitObject(Element PublicElement, Element PrivateElement, String OpFrom) {
/*
List nodelist;Element RFElement=null;Element node;int i=0;String Name="FunctionManager";
if ( PublicElement == null ) System.out.println("PublicElement=null");
if ( PrivateElement == null ) System.out.println("PrivateElement=null");
RFElement = JActiveDComDM.XMLRegistry.GetElementByName(PublicElement,Name);
if ( RFElement == null ) return;
InitFunctionConfigXML(RFElement.getAttribute("functionconfig").getValue());
JFunctionStub FS;
nodelist = JActiveDComDM.XMLRegistry.BeginEnumerate(RFElement);
while ( nodelist != null ) {
node = JActiveDComDM.XMLRegistry.Enumerate(nodelist,i);
if ( node == null ) break;
// if ( node.getNodeType() == node.ELEMENT_NODE ) {
RFElement = (Element)node;
FS = new JFunctionStub();
FS.FunctionID = RFElement.getAttribute("id").getValue().toUpperCase();
FS.ClassName = RFElement.getAttribute("class").getValue();
FunctionList.add(FS);
// }
i++;
}
JActiveDComDM.XMLRegistry.EndEnumerate();
*/
Element RFElement = null;
Element node;
int i = 0;
String Name = "FunctionManager";
if (PublicElement == null) System.out.println("PublicElement=null");
if (PrivateElement == null) System.out.println("PrivateElement=null");
RFElement = JActiveDComDM.XMLRegistry.GetElementByName(PublicElement, Name);
if (RFElement == null) return;
InitFunctionConfigXML(RFElement.getAttribute("functionconfig").getValue());
java.util.List nodelist = PackageStub.getContentVector("fmisReportFunction");
JFunctionStub FS;
StubObject so = null;
for (i = 0; i < nodelist.size(); i++) {
so = (StubObject) nodelist.get(i);
FS = new JFunctionStub();
FS.FunctionID = so.getString("id", "");
;
FS.ClassName = so.getString(OpFrom, "");
;
FunctionList.add(FS);
}
}
public String getPreviousStage(String currStage) {
if (currStage == null) {
return null;
}
int ind = currStage.indexOf(":");
Version v = new Version(currStage.substring(0, ind));
int lind = Integer.parseInt(currStage.substring(ind + 1));
java.util.List<WhatsNewItem> its = this.items.get(v);
if (its == null) {
return null;
}
lind--;
if (lind > -1) {
return v.getVersion() + ":" + lind;
}
Version p = this.items.lowerKey(v);
if (p == null) {
return null;
}
java.util.List<WhatsNewItem> pits = this.items.get(p);
if (pits != null) {
return p.getVersion() + ":" + (pits.size() - 1);
}
return null;
}
public String getNextStage(String currStage) {
if (currStage == null) {
return this.getStartStage();
}
int ind = currStage.indexOf(":");
Version v = new Version(currStage.substring(0, ind));
int lind = Integer.parseInt(currStage.substring(ind + 1));
java.util.List<WhatsNewItem> its = this.items.get(v);
if (its == null) {
return null;
}
lind++;
if (lind <= (its.size() - 1)) {
return v.getVersion() + ":" + lind;
}
Version n = this.items.higherKey(v);
if (n == null) {
return null;
}
java.util.List<WhatsNewItem> nits = this.items.get(n);
if (nits != null) {
return n.getVersion() + ":0";
}
return null;
}
/**
* 设置默认值
*
* @param dv
* @param formName
* @param actionType
* @return
*/
private S_FrameWork_Query_OperationValueObject setDefault(
S_FrameWork_Query_OperationValueObject valueObject, String formName, String actionType) {
String entityName = "S_FrameWork_Query_Operation"; // 代码自动生成
Document doc = DomService.getXMLDocFromEntity(entityName, formName);
// 1.遍历所有节点，判断是否有默认设置
// 2.比较actionType，看是否当前默认设置的操作
// 3.反射设置当前属性的默认值
try {
if (doc == null) {
return valueObject;
}
String methordName = "";
Element root = doc.getRootElement();
java.util.List elements = root.getChildren("item");
Element tempElm = null;
String s = null;
String sType = "string";
String sDBType = "String";
DefaultValue dv = new DefaultValue();
String whendefault = "";
String defaultValue = "";
for (int i = 0; i < elements.size(); i++) {
tempElm = (Element) elements.get(i);
s = tempElm.getAttribute("name").getValue();
methordName = "set" + s.substring(0, 1).toUpperCase() + s.substring(1, s.length());
if ((tempElm.getAttribute("datatype") != null
&& tempElm.getAttribute("datatype").getValue() != null)) {
if (tempElm.getAttribute("dbtype") == null) {
sType = tempElm.getAttribute("datatype").getValue();
} else {
sType = tempElm.getAttribute("dbtype").getValue();
}
}
// 设置缺省值;当新增加的时候，设置缺省值，此处只在初始化和load的时候设置；修改的时候不设置
// updated by wzp 20050323
if (tempElm.getAttribute("datatype") != null
&& tempElm.getAttribute("default") != null
&& tempElm.getAttribute("whendefault") != null) {
whendefault = tempElm.getAttribute("whendefault").getValue();
defaultValue = tempElm.getAttribute("default").getValue();
// 如何判断是新增？
if (actionType.toLowerCase().indexOf("init_") == 0
&& whendefault.toLowerCase().indexOf("init") > -1) {
// 初始化
valueObject = this.invokeMethord(valueObject, methordName, sType, defaultValue);
} else if (actionType.toLowerCase().indexOf("create_") == 0
&& whendefault.toLowerCase().indexOf("create") > -1) {
// 创建
valueObject = this.invokeMethord(valueObject, methordName, sType, defaultValue);
} else if (actionType.toLowerCase().indexOf("save_") == 0
&& whendefault.toLowerCase().indexOf("save") > -1) {
// 修改
valueObject = this.invokeMethord(valueObject, methordName, sType, defaultValue);
}
}
}
} catch (Exception exe) {
exe.printStackTrace();
}
return valueObject;
}
/**
* Gets the SER for re-ranking predicted by the machine learning technique.
*
* @param xmlFileName the XML file name containing the reference data (i.e. the human utterance
* transcription).
* @param svmPredFileName - file name of the hypothetised data (i.e. the re-estimated recognition
* )
* @return Semantic Error Rate
* @throws Exception
*/
public float[] getER4Predicted(String xmlFileName, String svmPredFileName) throws Exception {
// read the xml file in order to get the utterance transcripts
try {
BufferedReader br = getBufferedReader(svmPredFileName);
Document d =
new org.jdom.input.SAXBuilder().build(new File(xmlFileName)); // PARSE THE XML FILE
java.util.List nbestList = d.getRootElement().getChildren("nbest_data");
float[] serArray = new float[nbestList.size()];
int noUtt = 0;
for (int i = 0; i < nbestList.size(); i++) {
Element nbestElem = (Element) nbestList.get(i);
noUtt++;
// In the xml tree: find utterance features for the 1-best predicted hypothesis
java.util.List recList = nbestElem.getChildren("recognition");
// Find the hypothesis that is classified by re-ranking procedure as minimal in rank:
float bestRank = 100000;
int bestRankIdx = -1;
// Take from predictions file: the index of the 1-best
for (int k = 1; k < recList.size(); k++) {
String thisLine = br.readLine();
if (thisLine == null)
throw new Exception(
"End of file in predictions file before reading the entire xml file!");
float predictedRank = new Float(thisLine).floatValue();
if (bestRank > predictedRank) {
bestRankIdx = k;
bestRank = predictedRank;
}
}
// bestRankIdx = 1; // just for checking what's the SER for always choosing first hyp as the
// best
// System.out.println("best rank = " + bestRankIdx);
if (recList.size() < bestRankIdx)
System.out.println("Less than " + bestRankIdx + " recognitions !!! ");
Element recElem = (Element) recList.get(bestRankIdx);
Element semCorrectElem = recElem.getChild("semantically_correct");
String semCorrect = semCorrectElem.getValue();
if (semCorrect.equalsIgnoreCase("good")) serArray[i] = 0;
else serArray[i] = 1;
}
String thisLine;
if ((thisLine = br.readLine()) != null) {
System.out.println(
"There are still some " + "predictions while the xml file reached endOfFile!!! ");
System.out.println(thisLine);
}
return serArray;
} catch (IOException eIO) {
eIO.printStackTrace();
} catch (JDOMException eJDOM) {
eJDOM.printStackTrace();
} catch (NullPointerException nullE) {
nullE.printStackTrace();
}
return null;
}
/**
* Gets the SER for linear re-ranking. Two main steps are performed: * 1) Performs re-ranking
* using a linear function, which takes into account ..., i.e. newRank = ... <br>
* 2) Takes the hypothesis which is smallest in (new) rank as the best one. <br>
* 3) Computes the semantic error rate (SER) as defined in the "semantically_correct" tag
*/
public float[] getER4LinearReranking(String xmlFileName) throws Exception {
// read the xml file in order to get the utterance transcripts
try {
Document d =
new org.jdom.input.SAXBuilder().build(new File(xmlFileName)); // PARSE THE XML FILE
java.util.List nbestList = d.getRootElement().getChildren("nbest_data");
float[] serArray = new float[nbestList.size()];
int noUtt = 0;
int minNewRankID = 1;
for (int i = 0; i < nbestList.size(); i++) {
Element nbestElem = (Element) nbestList.get(i);
noUtt++;
java.util.List recList = nbestElem.getChildren("recognition");
///////////////////////////////////////////
int minNewRank = 100;
for (int j = 1; j < recList.size(); j++) {
Element recElem = (Element) recList.get(j);
int rank = new Integer(recElem.getChild("rank").getValue()).intValue();
int no_dialogue_move =
new Integer(recElem.getChild("no_dialogue_move").getValue()).intValue();
int underconstrained_query =
new Integer(recElem.getChild("underconstrained_query").getValue()).intValue();
int non_indefinite_existential =
new Integer(recElem.getChild("non_indefinite_existential").getValue()).intValue();
int non_show_imperative =
new Integer(recElem.getChild("non_show_imperative").getValue()).intValue();
int indefinite_meeting_and_meeting_referent =
new Integer(recElem.getChild("indefinite_meeting_and_meeting_referent").getValue())
.intValue();
/*
% 1) Place in the N-best list - lower number is better
feature_weight(rank, -1).
% 2) Strongly penalise hypotheses that produce no dialogue move
feature_weight(no_dialogue_move, -50).
% 3) Penalise queries with no contentful constraints
feature_weight(underconstrained_query, -10).
% 4) Penalise existentials which aren't indefinite, e.g. "is there the meeting next week"
feature_weight(non_indefinite_existential, -10).
% 5) Strongly penalise imperatives where the main verb isn't "show" or something similar
feature_weight(non_show_imperative, -50).
% 6) Disprefer combination of indefinite mention of meeting + available meeting referent
feature_weight(indefinite_meeting_and_meeting_referent, -2).
*/
int newRank =
rank
+ 50 * no_dialogue_move
+ 10 * underconstrained_query
+ 10 * non_indefinite_existential
+ 50 * non_show_imperative
+ 2 * indefinite_meeting_and_meeting_referent;
if (newRank < minNewRank) {
minNewRank = newRank;
minNewRankID = j;
}
}
Element recElem = (Element) recList.get(minNewRankID);
//////////////////////////////////////////
// System.out.println("best rank = " + minNewRankID);
if (recList.size() < minNewRankID)
System.out.println("Less than " + minNewRankID + " recognitions !!! ");
Element semCorrectElem = recElem.getChild("semantically_correct");
String semCorrect = semCorrectElem.getValue();
if (semCorrect.equalsIgnoreCase("good")) serArray[i] = 0;
else serArray[i] = 1;
}
return serArray;
} catch (IOException eIO) {
eIO.printStackTrace();
} catch (JDOMException eJDOM) {
eJDOM.printStackTrace();
} catch (NullPointerException nullE) {
nullE.printStackTrace();
}
return null;
}
public WhatsNew(AbstractProjectViewer pv, boolean onlyShowCurrentVersion)
throws GeneralException {
super(pv);
String wn = Environment.getProperty(Constants.WHATS_NEW_VERSION_VIEWED_PROPERTY_NAME);
if (wn == null) {
wn = "0";
}
// Get the current whats new version (i.e. old).
Version lastWhatsNewVersion = new Version(wn);
boolean betasAllowed =
Environment.getUserProperties()
.getPropertyAsBoolean(Constants.OPTIN_TO_BETA_VERSIONS_PROPERTY_NAME);
try {
String whatsNew = Environment.getResourceFileAsString(Constants.WHATS_NEW_FILE);
// Load up all the whats new for greater versions.
Element root = JDOMUtils.getStringAsElement(whatsNew);
java.util.List verEls = JDOMUtils.getChildElements(root, XMLConstants.version, false);
// Assume they are in the right order
// TODO: Enforce the order and/or sort.
for (int i = 0; i < verEls.size(); i++) {
Element vEl = (Element) verEls.get(i);
String id = JDOMUtils.getAttributeValue(vEl, XMLConstants.id, true);
Version v = new Version(id);
/*
if ((v.isBeta ())
&&
(!betasAllowed)
)
{
// Ignore, the user isn't interested in betas.
continue;
}
*/
if ((lastWhatsNewVersion.isNewer(v))
|| ((onlyShowCurrentVersion) && (v.isSame(Environment.getQuollWriterVersion())))) {
String c = WhatsNewComponentProvider.class.getName();
int ind = c.lastIndexOf(".");
if (ind > 0) {
c = c.substring(0, ind);
}
WhatsNewComponentProvider compProv = null;
String cl = JDOMUtils.getAttributeValue(vEl, XMLConstants.clazz, false);
if (!cl.equals("")) {
Class clz = null;
try {
clz = Class.forName(cl);
if (WhatsNewComponentProvider.class.isAssignableFrom(clz)) {
compProv = (WhatsNewComponentProvider) clz.newInstance();
}
} catch (Exception e) {
}
}
// This is a version we are interested in.
java.util.List itemEls =
JDOMUtils.getChildElements(vEl, WhatsNewItem.XMLConstants.root, true);
java.util.List<WhatsNewItem> its = new ArrayList();
for (int j = 0; j < itemEls.size(); j++) {
Element itEl = (Element) itemEls.get(j);
WhatsNewItem it = new WhatsNewItem(itEl, compProv, pv);
if (it.onlyIfCurrentVersion) {
if (!Environment.getQuollWriterVersion().isSame(v)) {
continue;
}
}
if ((it.description == null) && (it.component == null)) {
Environment.logMessage(
"Whats new item has no description or component, referenced by: "
+ JDOMUtils.getPath(itEl));
continue;
}
its.add(it);
}
if (its.size() > 0) {
this.items.put(v, its);
}
}
}
} catch (Exception e) {
throw new GeneralException("Unable to init whats new", e);
}
}