/* Copyright (C) 2002 Univ. of Massachusetts Amherst, Computer Science Dept. This file is part of "MALLET" (MAchine Learning for LanguagE Toolkit). http://www.cs.umass.edu/~mccallum/mallet This software is provided under the terms of the Common Public License, version 1.0, as published by http://www.opensource.org. For further information, see the file `LICENSE' included with this distribution. */ /** Evaluate segmentation f1 for several different tags (marked in OIB format). For example, tags might be B-PERSON I-PERSON O B-LOCATION I-LOCATION O... @author Andrew McCallum <a href="mailto:mccallum@cs.umass.edu">mccallum@cs.umass.edu</a> */ package cc.mallet.fst; import java.io.PrintStream; import java.util.List; import java.util.logging.Logger; import java.text.DecimalFormat; import cc.mallet.types.FeatureVector; import cc.mallet.types.Instance; import cc.mallet.types.InstanceList; import cc.mallet.types.Sequence; import cc.mallet.types.TokenSequence; import cc.mallet.util.MalletLogger; /** * Evaluates a transducer model, computes the precision, recall and F1 scores; * considers segments that span across multiple tokens. */ public class MultiSegmentationEvaluator extends TransducerEvaluator { private static Logger logger = MalletLogger.getLogger(SegmentationEvaluator.class.getName()); // equals() is called on these objects to determine if this token is the start or continuation of a segment. // A tag not equal to any of these is an "other". // is not part of the segment). Object[] segmentStartTags; Object[] segmentContinueTags; Object[] segmentStartOrContinueTags; public MultiSegmentationEvaluator (InstanceList[] instanceLists, String[] instanceListDescriptions, Object[] segmentStartTags, Object[] segmentContinueTags) { super (instanceLists, instanceListDescriptions); this.segmentStartTags = segmentStartTags; this.segmentContinueTags = segmentContinueTags; assert (segmentStartTags.length == segmentContinueTags.length); } public MultiSegmentationEvaluator (InstanceList instanceList1, String description1, Object[] segmentStartTags, Object[] segmentContinueTags) { this (new InstanceList[] {instanceList1}, new String[] {description1}, segmentStartTags, segmentContinueTags); } public MultiSegmentationEvaluator (InstanceList instanceList1, String description1, InstanceList instanceList2, String description2, Object[] segmentStartTags, Object[] segmentContinueTags) { this (new InstanceList[] {instanceList1, instanceList2}, new String[] {description1, description2}, segmentStartTags, segmentContinueTags); } public MultiSegmentationEvaluator (InstanceList instanceList1, String description1, InstanceList instanceList2, String description2, InstanceList instanceList3, String description3, Object[] segmentStartTags, Object[] segmentContinueTags) { this (new InstanceList[] {instanceList1, instanceList2, instanceList3}, new String[] {description1, description2, description3}, segmentStartTags, segmentContinueTags); } public void evaluateInstanceList (TransducerTrainer tt, InstanceList data, String description) { Transducer model = tt.getTransducer(); int numCorrectTokens, totalTokens; int[] numTrueSegments, numPredictedSegments, numCorrectSegments; int allIndex = segmentStartTags.length; numTrueSegments = new int[allIndex+1]; numPredictedSegments = new int[allIndex+1]; numCorrectSegments = new int[allIndex+1]; totalTokens = numCorrectTokens = 0; for (int n = 0; n < numTrueSegments.length; n++) numTrueSegments[n] = numPredictedSegments[n] = numCorrectSegments[n] = 0; for (int i = 0; i < data.size(); i++) { Instance instance = data.get(i); Sequence input = (Sequence) instance.getData(); //String tokens = null; //if (instance.getSource() != null) //tokens = (String) instance.getSource().toString(); Sequence trueOutput = (Sequence) instance.getTarget(); assert (input.size() == trueOutput.size()); Sequence predOutput = model.transduce (input); assert (predOutput.size() == trueOutput.size()); int trueStart, predStart; // -1 for non-start, otherwise index into segmentStartTag for (int j = 0; j < trueOutput.size(); j++) { totalTokens++; if (trueOutput.get(j).equals(predOutput.get(j))) numCorrectTokens++; trueStart = predStart = -1; // Count true segment starts for (int n = 0; n < segmentStartTags.length; n++) { if (segmentStartTags[n].equals(trueOutput.get(j))) { numTrueSegments[n]++; numTrueSegments[allIndex]++; trueStart = n; break; } } // Count predicted segment starts for (int n = 0; n < segmentStartTags.length; n++) { if (segmentStartTags[n].equals(predOutput.get(j))) { numPredictedSegments[n]++; numPredictedSegments[allIndex]++; predStart = n; } } if (trueStart != -1 && trueStart == predStart) { // Truth and Prediction both agree that the same segment tag-type is starting now int m; boolean trueContinue = false; boolean predContinue = false; for (m = j+1; m < trueOutput.size(); m++) { trueContinue = segmentContinueTags[predStart].equals (trueOutput.get(m)); predContinue = segmentContinueTags[predStart].equals (predOutput.get(m)); if (!trueContinue || !predContinue) { if (trueContinue == predContinue) { // They agree about a segment is ending somehow numCorrectSegments[predStart]++; numCorrectSegments[allIndex]++; } break; } } // for the case of the end of the sequence if (m == trueOutput.size()) { if (trueContinue == predContinue) { numCorrectSegments[predStart]++; numCorrectSegments[allIndex]++; } } } } } DecimalFormat f = new DecimalFormat ("0.####"); logger.info (description +" tokenaccuracy="+f.format(((double)numCorrectTokens)/totalTokens)); for (int n = 0; n < numCorrectSegments.length; n++) { logger.info ((n < allIndex ? segmentStartTags[n].toString() : "OVERALL") +' '); double precision = numPredictedSegments[n] == 0 ? 1 : ((double)numCorrectSegments[n]) / numPredictedSegments[n]; double recall = numTrueSegments[n] == 0 ? 1 : ((double)numCorrectSegments[n]) / numTrueSegments[n]; double f1 = recall+precision == 0.0 ? 0.0 : (2.0 * recall * precision) / (recall + precision); logger.info (" "+description+" segments true="+numTrueSegments[n]+" pred="+numPredictedSegments[n]+" correct="+numCorrectSegments[n]+ " misses="+(numTrueSegments[n]-numCorrectSegments[n])+" alarms="+(numPredictedSegments[n]-numCorrectSegments[n])); logger.info (" "+description+" precision="+f.format(precision)+" recall="+f.format(recall)+" f1="+f.format(f1)); } } /** * Returns the number of incorrect segments in <code>predOutput</code> * * @param trueOutput truth * @param predOutput predicted * @return number of incorrect segments */ public int numIncorrectSegments (Sequence trueOutput, Sequence predOutput) { int numCorrectTokens, totalTokens; int[] numTrueSegments, numPredictedSegments, numCorrectSegments; int allIndex = segmentStartTags.length; numTrueSegments = new int[allIndex+1]; numPredictedSegments = new int[allIndex+1]; numCorrectSegments = new int[allIndex+1]; totalTokens = numCorrectTokens = 0; for (int n = 0; n < numTrueSegments.length; n++) numTrueSegments[n] = numPredictedSegments[n] = numCorrectSegments[n] = 0; assert (predOutput.size() == trueOutput.size()); // -1 for non-start, otherwise index into segmentStartTag int trueStart, predStart; for (int j = 0; j < trueOutput.size(); j++) { totalTokens++; if (trueOutput.get(j).equals(predOutput.get(j))) numCorrectTokens++; trueStart = predStart = -1; // Count true segment starts for (int n = 0; n < segmentStartTags.length; n++) { if (segmentStartTags[n].equals(trueOutput.get(j))) { numTrueSegments[n]++; numTrueSegments[allIndex]++; trueStart = n; break; } } // Count predicted segment starts for (int n = 0; n < segmentStartTags.length; n++) { if (segmentStartTags[n].equals(predOutput.get(j))) { numPredictedSegments[n]++; numPredictedSegments[allIndex]++; predStart = n; } } if (trueStart != -1 && trueStart == predStart) { // Truth and Prediction both agree that the same segment tag-type is starting now int m; boolean trueContinue = false; boolean predContinue = false; for (m = j+1; m < trueOutput.size(); m++) { trueContinue = segmentContinueTags[predStart].equals (trueOutput.get(m)); predContinue = segmentContinueTags[predStart].equals (predOutput.get(m)); if (!trueContinue || !predContinue) { if (trueContinue == predContinue) { // They agree about a segment is ending somehow numCorrectSegments[predStart]++; numCorrectSegments[allIndex]++; } break; } } // for the case of the end of the sequence if (m == trueOutput.size()) { if (trueContinue == predContinue) { numCorrectSegments[predStart]++; numCorrectSegments[allIndex]++; } } } } int wrong = 0; for (int n=0; n < numCorrectSegments.length; n++) { // incorrect segment is either false pos or false neg. wrong += numTrueSegments[n] - numCorrectSegments[n]; } return wrong; } /** * Tests segmentation using an ArrayList of predicted Sequences instead of a * {@link Transducer}. If predictedSequence is null, don't include in stats * (useful for error analysis). * * @param data list of instances to be segmented * @param predictedSequences predictions * @param description description of trial * @param viterbiOutputStream where to print the Viterbi paths */ public void batchTest(InstanceList data, List<Sequence> predictedSequences, String description, PrintStream viterbiOutputStream) { int numCorrectTokens, totalTokens; int[] numTrueSegments, numPredictedSegments, numCorrectSegments; int allIndex = segmentStartTags.length; numTrueSegments = new int[allIndex+1]; numPredictedSegments = new int[allIndex+1]; numCorrectSegments = new int[allIndex+1]; TokenSequence sourceTokenSequence = null; totalTokens = numCorrectTokens = 0; for (int n = 0; n < numTrueSegments.length; n++) numTrueSegments[n] = numPredictedSegments[n] = numCorrectSegments[n] = 0; for (int i = 0; i < data.size(); i++) { if (viterbiOutputStream != null) viterbiOutputStream.println ("Viterbi path for "+description+" instance #"+i); Instance instance = data.get(i); Sequence input = (Sequence) instance.getData(); //String tokens = null; //if (instance.getSource() != null) //tokens = (String) instance.getSource().toString(); Sequence trueOutput = (Sequence) instance.getTarget(); assert (input.size() == trueOutput.size()); Sequence predOutput = (Sequence) predictedSequences.get (i); if (predOutput == null) // skip this instance continue; assert (predOutput.size() == trueOutput.size()); int trueStart, predStart; // -1 for non-start, otherwise index into segmentStartTag for (int j = 0; j < trueOutput.size(); j++) { totalTokens++; if (trueOutput.get(j).equals(predOutput.get(j))) numCorrectTokens++; trueStart = predStart = -1; // Count true segment starts for (int n = 0; n < segmentStartTags.length; n++) { if (segmentStartTags[n].equals(trueOutput.get(j))) { numTrueSegments[n]++; numTrueSegments[allIndex]++; trueStart = n; break; } } // Count predicted segment starts for (int n = 0; n < segmentStartTags.length; n++) { if (segmentStartTags[n].equals(predOutput.get(j))) { numPredictedSegments[n]++; numPredictedSegments[allIndex]++; predStart = n; } } if (trueStart != -1 && trueStart == predStart) { // Truth and Prediction both agree that the same segment tag-type is starting now int m; boolean trueContinue = false; boolean predContinue = false; for (m = j+1; m < trueOutput.size(); m++) { trueContinue = segmentContinueTags[predStart].equals (trueOutput.get(m)); predContinue = segmentContinueTags[predStart].equals (predOutput.get(m)); if (!trueContinue || !predContinue) { if (trueContinue == predContinue) { // They agree about a segment is ending somehow numCorrectSegments[predStart]++; numCorrectSegments[allIndex]++; } break; } } // for the case of the end of the sequence if (m == trueOutput.size()) { if (trueContinue == predContinue) { numCorrectSegments[predStart]++; numCorrectSegments[allIndex]++; } } } if (viterbiOutputStream != null) { FeatureVector fv = (FeatureVector) input.get(j); //viterbiOutputStream.println (tokens.charAt(j)+" "+trueOutput.get(j).toString()+ //'/'+predOutput.get(j).toString()+" "+ fv.toString(true)); if (sourceTokenSequence != null) viterbiOutputStream.print (sourceTokenSequence.get(j).getText()+": "); viterbiOutputStream.println (trueOutput.get(j).toString()+ '/'+predOutput.get(j).toString()+" "+ fv.toString(true)); } } } DecimalFormat f = new DecimalFormat ("0.####"); logger.info (description +" tokenaccuracy="+f.format(((double)numCorrectTokens)/totalTokens)); for (int n = 0; n < numCorrectSegments.length; n++) { logger.info ((n < allIndex ? segmentStartTags[n].toString() : "OVERALL") +' '); double precision = numPredictedSegments[n] == 0 ? 1 : ((double)numCorrectSegments[n]) / numPredictedSegments[n]; double recall = numTrueSegments[n] == 0 ? 1 : ((double)numCorrectSegments[n]) / numTrueSegments[n]; double f1 = recall+precision == 0.0 ? 0.0 : (2.0 * recall * precision) / (recall + precision); logger.info (" segments true="+numTrueSegments[n]+" pred="+numPredictedSegments[n]+" correct="+numCorrectSegments[n]+ " misses="+(numTrueSegments[n]-numCorrectSegments[n])+" alarms="+(numPredictedSegments[n]-numCorrectSegments[n])); logger.info (" precision="+f.format(precision)+" recall="+f.format(recall)+" f1="+f.format(f1)); } } }