Examples with CodingAnnotationStudy org.dkpro.statistics.agreement.coding.CodingAnnotationStudy used on opensource projects

Example 1 with CodingAnnotationStudy

use of org.dkpro.statistics.agreement.coding.CodingAnnotationStudy in project webanno by webanno.

the class AgreementUtils method makeCodingStudy.

private static CodingAgreementResult makeCodingStudy(CasDiff aDiff, Collection<String> aUsers, String aType, String aFeature, boolean aExcludeIncomplete, boolean aNullLabelsAsEmpty, Map<String, List<CAS>> aCasMap) {
    List<String> users = new ArrayList<>(aUsers);
    Collections.sort(users);
    List<ConfigurationSet> completeSets = new ArrayList<>();
    List<ConfigurationSet> setsWithDifferences = new ArrayList<>();
    List<ConfigurationSet> incompleteSetsByPosition = new ArrayList<>();
    List<ConfigurationSet> incompleteSetsByLabel = new ArrayList<>();
    List<ConfigurationSet> pluralitySets = new ArrayList<>();
    List<ConfigurationSet> irrelevantSets = new ArrayList<>();
    CodingAnnotationStudy study = new CodingAnnotationStudy(users.size());
    // Check if the feature we are looking at is a primitive feature or a link feature
    // We do this by looking it up in the first available CAS. Mind that at this point all
    // CASes should have exactly the same typesystem.
    CAS someCas = findSomeCas(aCasMap);
    if (someCas == null) {
        // Well... there is NOTHING here!
        // All positions are irrelevant
        aDiff.getPositions().forEach(p -> irrelevantSets.add(aDiff.getConfigurationSet(p)));
        return new CodingAgreementResult(aType, aFeature, aDiff.toResult(), study, users, completeSets, irrelevantSets, setsWithDifferences, incompleteSetsByPosition, incompleteSetsByLabel, pluralitySets, aExcludeIncomplete);
    }
    TypeSystem ts = someCas.getTypeSystem();
    // We should just do the right thing here which is: do nothing
    if (ts.getType(aType) == null) {
        // All positions are irrelevant
        aDiff.getPositions().forEach(p -> irrelevantSets.add(aDiff.getConfigurationSet(p)));
        return new CodingAgreementResult(aType, aFeature, aDiff.toResult(), study, users, completeSets, irrelevantSets, setsWithDifferences, incompleteSetsByPosition, incompleteSetsByLabel, pluralitySets, aExcludeIncomplete);
    }
    // Check that the feature really exists instead of just getting a NPE later
    if (ts.getType(aType).getFeatureByBaseName(aFeature) == null) {
        throw new IllegalArgumentException("Type [" + aType + "] has no feature called [" + aFeature + "]");
    }
    boolean isPrimitiveFeature = ts.getType(aType).getFeatureByBaseName(aFeature).getRange().isPrimitive();
    nextPosition: for (Position p : aDiff.getPositions()) {
        ConfigurationSet cfgSet = aDiff.getConfigurationSet(p);
        // Only calculate agreement for the given layer
        if (!cfgSet.getPosition().getType().equals(aType)) {
            // We don't even consider these as irrelevant, they are just filtered out
            continue;
        }
        // If the feature on a position is set, then it is a subposition
        boolean isSubPosition = p.getFeature() != null;
        // this is an inverted XOR!
        if (!(isPrimitiveFeature ^ isSubPosition)) {
            irrelevantSets.add(cfgSet);
            continue;
        }
        // feature
        if (isSubPosition && !aFeature.equals(cfgSet.getPosition().getFeature())) {
            irrelevantSets.add(cfgSet);
            continue nextPosition;
        }
        // If non of the current users has made any annotation at this position, then skip it
        if (users.stream().filter(u -> cfgSet.getCasGroupIds().contains(u)).count() == 0) {
            irrelevantSets.add(cfgSet);
            continue nextPosition;
        }
        Object[] values = new Object[users.size()];
        int i = 0;
        for (String user : users) {
            // this configuration set.
            if (!cfgSet.getCasGroupIds().contains(user)) {
                incompleteSetsByPosition.add(cfgSet);
                if (aExcludeIncomplete) {
                    // Record as incomplete
                    continue nextPosition;
                } else {
                    // Record as missing value
                    values[i] = null;
                    i++;
                    continue;
                }
            }
            // Make sure a single user didn't do multiple alternative annotations at a single
            // position. So there is currently no support for calculating agreement on stacking
            // annotations.
            List<Configuration> cfgs = cfgSet.getConfigurations(user);
            if (cfgs.size() > 1) {
                pluralitySets.add(cfgSet);
                continue nextPosition;
            }
            Configuration cfg = cfgs.get(0);
            // Check if source and/or targets of a relation are stacked
            if (cfg.getPosition() instanceof RelationPosition) {
                RelationPosition pos = (RelationPosition) cfg.getPosition();
                FeatureStructure arc = cfg.getFs(user, pos.getCasId(), aCasMap);
                RelationDiffAdapter adapter = (RelationDiffAdapter) aDiff.getTypeAdapters().get(pos.getType());
                // Check if the source of the relation is stacked
                AnnotationFS source = FSUtil.getFeature(arc, adapter.getSourceFeature(), AnnotationFS.class);
                List<AnnotationFS> sourceCandidates = CasUtil.selectAt(arc.getCAS(), source.getType(), source.getBegin(), source.getEnd());
                if (sourceCandidates.size() > 1) {
                    pluralitySets.add(cfgSet);
                    continue nextPosition;
                }
                // Check if the target of the relation is stacked
                AnnotationFS target = FSUtil.getFeature(arc, adapter.getTargetFeature(), AnnotationFS.class);
                List<AnnotationFS> targetCandidates = CasUtil.selectAt(arc.getCAS(), target.getType(), target.getBegin(), target.getEnd());
                if (targetCandidates.size() > 1) {
                    pluralitySets.add(cfgSet);
                    continue nextPosition;
                }
            }
            // Only calculate agreement for the given feature
            FeatureStructure fs = cfg.getFs(user, cfg.getPosition().getCasId(), aCasMap);
            values[i] = extractValueForAgreement(fs, aFeature, cfg.getAID(user).index, cfg.getPosition().getLinkCompareBehavior());
            // agreement calculation. The empty label is still a valid label.
            if (aNullLabelsAsEmpty && values[i] == null) {
                values[i] = "";
            }
            // "null" cannot be used in agreement calculations. We treat these as incomplete
            if (values[i] == null) {
                incompleteSetsByLabel.add(cfgSet);
                if (aExcludeIncomplete) {
                    continue nextPosition;
                }
            }
            i++;
        }
        if (ObjectUtils.notEqual(values[0], values[1])) {
            setsWithDifferences.add(cfgSet);
        }
        // are calculating agreement over
        assert cfgSet.getPosition().getFeature() == null || cfgSet.getPosition().getFeature().equals(aFeature);
        completeSets.add(cfgSet);
        study.addItemAsArray(values);
    }
    return new CodingAgreementResult(aType, aFeature, aDiff.toResult(), study, users, completeSets, irrelevantSets, setsWithDifferences, incompleteSetsByPosition, incompleteSetsByLabel, pluralitySets, aExcludeIncomplete);
}