OpenNLP Maxent miscalculates for real values < 1
------------------------------------------------
Key: OPENNLP-170
URL: https://issues.apache.org/jira/browse/OPENNLP-170
Project: OpenNLP
Issue Type: Bug
Components: Maxent
Affects Versions: maxent-3.0.0-sourceforge
Environment: Windows 7, Java 1.6
Reporter: Assaf Urieli
When using predicates with real values, entering real values predA=0.1
predB=0.2 gives different results than predA=10, predB=20
However, using predA=1, predB=2 gives the same results as predA=10, predB=20.
Test below:
package openMaxentTest;
import java.io.StringReader;
import junit.framework.TestCase;
import opennlp.maxent.GIS;
import opennlp.maxent.PlainTextByLineDataStream;
import opennlp.maxent.RealBasicEventStream;
import opennlp.model.EventStream;
import opennlp.model.MaxentModel;
import opennlp.model.OnePassRealValueDataIndexer;
import opennlp.model.RealValueFileEventStream;
public class ScaleDoesntMatterTest extends TestCase {
/**
* This test sets out to prove that the scale you use on real valued
predicates
* doesn't matter when it comes the probability assigned to each
outcome.
* Strangely, if we use (1,2) and (10,20) there's no difference.
* If we use (0.1,0.2) and (10,20) there is a difference.
* @throws Exception
*/
public void testScaleResults() throws Exception {
String smallValues = "predA=0.1 predB=0.2 A\n" +
"predB=0.3 predA=0.1 B\n";
String smallTest = "predA=0.2 predB=0.2";
String largeValues = "predA=10 predB=20 A\n" +
"predB=30 predA=10 B\n";
String largeTest = "predA=20 predB=20";
StringReader smallReader = new StringReader(smallValues);
EventStream smallEventStream = new RealBasicEventStream(new
PlainTextByLineDataStream(smallReader));
MaxentModel smallModel = GIS.trainModel(2, new
OnePassRealValueDataIndexer(smallEventStream,0), false);
String[] contexts = smallTest.split(" ");
float[] values =
RealValueFileEventStream.parseContexts(contexts);
double[] ocs = smallModel.eval(contexts, values);
String smallResults = smallModel.getAllOutcomes(ocs);
System.out.println("smallResults: " + smallResults);
StringReader largeReader = new StringReader(largeValues);
EventStream largeEventStream = new RealBasicEventStream(new
PlainTextByLineDataStream(largeReader));
MaxentModel largeModel = GIS.trainModel(2, new
OnePassRealValueDataIndexer(largeEventStream,0), false);
contexts = largeTest.split(" ");
values = RealValueFileEventStream.parseContexts(contexts);
ocs = largeModel.eval(contexts, values);
String largeResults = smallModel.getAllOutcomes(ocs);
System.out.println("largeResults: " + largeResults);
assertEquals(smallResults, largeResults);
}
}
The problem concerns the correctionConstant in GISTrainer, which is set to be
an integer. I implemented the following fix in class GISTrainer:
// determine the correction constant and its inverse
//int correctionConstant = 1;
float correctionConstant = 0;
for (int ci = 0; ci < contexts.length; ci++) {
if (values == null || values[ci] == null) {
if (contexts[ci].length > correctionConstant) {
correctionConstant = contexts[ci].length;
}
}
else {
float cl = values[ci][0];
for (int vi=1;vi<values[ci].length;vi++) {
cl+=values[ci][vi];
}
if (cl > correctionConstant) {
//correctionConstant=(int) Math.ceil(cl);
correctionConstant= cl;
}
}
}
I'd be curious to know if there's a reason for using an integer
correctionConstant.
Rgds,
Assaf Urieli
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
