shalmaneser-rosy 1.2.0.rc4

data/lib/rosy/RosySplit.rb

@@ -0,0 +1,232 @@
+# RosySplit
+# KE, SP May 05
+#
+# One of the main task modules of Rosy:
+# split training data into training and test parts
+#
+# A split is realized as two DB tables, 
+# one with the sentence IDs of the training part of the split,
+# and one with the sentence IDs of the test part of the split.
+#
+# Additionally, each split table also contains all phase-2 features
+# for the train/test part of the split:
+# Phase 2 features are trained on training features and applied to
+# test features. They need to be retrained for each split.
+
+require "common/ruby_class_extensions"
+
+# Frprep packages
+require "common/prep_config_data"
+
+# Rosy packages
+require "rosy/FailedParses"
+require "rosy/FeatureInfo"
+require "common/RosyConventions"
+require "rosy/RosyIterator"
+require "rosy/RosyTask"
+require "rosy/RosyTrainingTestTable"
+require "rosy/View"
+
+class RosySplit < RosyTask
+
+  def initialize(exp,      # RosyConfigData object: experiment description
+		 opts,     # hash: runtime argument option (string) -> value (string)
+		 ttt_obj)  # RosyTrainingTestTable object
+
+    #####
+    # In enduser mode, this whole task is unavailable
+    in_enduser_mode_unavailable()
+
+    ##
+    # remember the experiment description
+
+    @exp = exp
+    @ttt_obj = ttt_obj
+
+
+    ##
+    # check runtime options
+
+    # default values
+    @trainpercent = 90
+    @splitID = nil
+
+    opts.each do |opt,arg|
+      case opt
+      when "--trainpercent"
+        @trainpercent = arg.to_i
+      when "--logID"
+        @splitID = arg
+      else
+	# this is an option that is okay but has already been read and used by rosy.rb
+      end	
+    end
+
+    #sanity checks
+    if @splitID.nil?
+      raise "I need an ID for the split in order to proceed. Parameter: --logID|-l"
+    end
+    if @trainpercent <= 0 or @trainpercent >= 100
+      raise "--trainpercent must be between 1 and 99."
+    end
+
+    # add preprocessing information to the experiment file object
+    # so we know what language the training data is in
+    preproc_filename = @exp.get("preproc_descr_file_train")
+    if not(preproc_filename)
+      $stderr.puts "Please set the name of the preprocessing exp. file name"
+      $stderr.puts "in the experiment file, parameter preproc_descr_file_train."
+      exit 1
+    elsif not(File.readable?(preproc_filename))
+      $stderr.puts "Error in the experiment file:"
+      $stderr.puts "Parameter preproc_descr_file_train has to be a readable file."
+      exit 1
+    end
+    preproc_exp = FrPrepConfigData.new(preproc_filename)
+    @exp.adjoin(preproc_exp)
+
+    # announce the task
+    $stderr.puts "---------"
+    $stderr.puts "Rosy experiment #{@exp.get("experiment_ID")}: Making split with ID #{@splitID}, training data percentage #{@trainpercent}%"
+    $stderr.puts "---------"
+  end
+
+  #####
+  # perform
+  #
+  # perform a split of the training data and the "failed sentences" object 
+  # the split is written to a DB table, the failed sentence splits are written to files 
+  def perform()
+
+    #################################
+    # 1. treat the failed sentences
+    perform_failed_parses()    
+    
+    ###############################
+    # 2. get the main table, split it, and write the result to two new tables
+    perform_make_split()
+
+    ###############################
+    # 3. Repeat the training and extraction of phase 2 features for this split,
+    #    and write the result to the split tables
+
+  end
+  
+  #######
+  # split index column name
+  def RosySplit.split_index_colname()
+    return "split_index"
+  end
+
+  ############
+  # make_join_restriction
+  #
+  # Given a splitID, the main table to be split, 
+  # the dataset (train or test), and the experiment file object,
+  # make a ValueRestriction object that can be passed to a view initialization:
+  #
+  # restrict main table rows to those that occur in the correct part
+  # (part = train or part = test) of the split with the given ID
+  #
+  # returns: VarVarRestriction object
+  def RosySplit.make_join_restriction(splitID,  # string: splitlogID
+				      table,    # DBtable object
+				      dataset,  # string: "train", "test"
+				      ttt_obj)  # RosyTrainingTestTable object
+
+    return VarVarRestriction.new(table.table_name + "." + table.index_name,
+                                 ttt_obj.splittable_name(splitID, dataset) + "." + RosySplit.split_index_colname())
+      
+  end
+
+  ###########
+  private
+
+  ##########
+  # perform_failed_parses:
+  #
+  # this is the part of the perform() method
+  # that splits the sentences with failed parses
+  # into a training and a test part
+  # and remembers this split
+  def perform_failed_parses()
+    # read file with failed parses    
+    failed_parses_filename = 
+          File.new_filename(@exp.instantiate("rosy_dir",
+                                             "exp_ID" => @exp.get("experiment_ID")),
+                            @exp.instantiate("failed_file",
+                                             "exp_ID" => @exp.get("experiment_ID"),
+                                             "split_ID" => "none",
+                                             "dataset" => "none"))
+    
+
+    fp_obj = FailedParses.new()
+    fp_obj.load(failed_parses_filename)
+
+    # split and write to appropriate files
+    fp_train_obj, fp_test_obj = fp_obj.make_split(@trainpercent)
+    
+    train_filename = 
+         File.new_filename(@exp.instantiate("rosy_dir",
+                                            "exp_ID" => @exp.get("experiment_ID")),
+                           @exp.instantiate("failed_file",
+                                            "exp_ID" => @exp.get("experiment_ID"),
+                                            "split_ID" => @splitID,
+                                            "dataset" => "train"))
+    
+    fp_train_obj.save(train_filename)
+    
+    test_filename = 
+        File.new_filename(@exp.instantiate("rosy_dir",
+                                           "exp_ID" => @exp.get("experiment_ID")),
+                          @exp.instantiate("failed_file",
+                                           "exp_ID" => @exp.get("experiment_ID"),
+                                           "split_ID" => @splitID,
+                                           "dataset" => "test"))
+
+    fp_test_obj.save(test_filename)
+  end
+
+  ##########
+  # perform_make_split
+  #
+  # this is the part of the perform() method
+  # that makes the actual split
+  # at random and stores it in new database tables
+  def perform_make_split()
+    $stderr.puts "Making split with ID #{@splitID}"
+
+    # get a view of the main table
+    maintable = @ttt_obj.existing_train_table()
+
+    # construct new DB tables for the train and test part of the new split:
+    # get table name and join column name 
+    split_train_table = @ttt_obj.new_split_table(@splitID, "train", RosySplit.split_index_colname())
+    split_test_table =  @ttt_obj.new_split_table(@splitID, "test", RosySplit.split_index_colname())
+    
+    # make split: put each sentence ID into either the train or the test table
+    # based on whether a random number btw. 0 and 100 is larger than @trainpercent or not
+    
+    
+    # go through training data one frame at a time
+    iterator = RosyIterator.new(@ttt_obj, @exp, "train", "xwise"=>"frame")
+    iterator.each_group { |dummy1, dummy2|
+      view = iterator.get_a_view_for_current_group(["sentid", maintable.index_name])
+      view.each_sentence() { |sentence|
+        if rand(100) > @trainpercent
+          # put this sentence into the test table
+          table = split_test_table
+        else
+          # put this sentence into the training table
+          table = split_train_table
+        end
+        sentence.each { |instance|
+          table.insert_row([[RosySplit.split_index_colname(), instance[maintable.index_name]],
+                            ["sentid", instance["sentid"]]])
+        }
+      }    
+      view.close()
+    }
+  end
+
+end

data/lib/rosy/RosyTask.rb

@@ -0,0 +1,19 @@
+##
+# RosyTask
+# KE, SP April 05
+#
+# this is the abstract class that describes the interface for
+# the task classes of Rosy. 
+#
+# all task classes should have a perform() method that actually
+# performs the task.
+
+class RosyTask
+  def initialize()
+    raise "Shouldn't be here! I'm an abstract class"
+  end
+
+  def perform()
+    raise "Should be overwritten by the inheriting class!"
+  end
+end

data/lib/rosy/RosyTest.rb

@@ -0,0 +1,829 @@
+# RosyTest
+# KE May 05
+#
+# One of the main task modules of Rosy:
+# apply classifiers
+
+# Standard library packages
+require "tempfile"
+require 'fileutils'
+
+# Salsa packages
+require "common/Parser"
+require "common/SalsaTigerRegXML"
+require "common/SynInterfaces"
+require "common/ruby_class_extensions"
+
+# Rosy packages
+require "rosy/FeatureInfo"
+require "common/ML"
+require "common/RosyConventions"
+require "rosy/RosyIterator"
+require "rosy/RosyTask"
+require "rosy/RosyTrainingTestTable"
+require "rosy/View"
+
+# Frprep packages
+#require "common/prep_config_data" # AB: what the fuck???
+
+##########################################################################
+# classifier combination class
+class ClassifierCombination
+
+  # new(): just remember experiment file object
+  def initialize(exp)
+    @exp = exp
+  end
+
+  # combine:
+  #
+  # given a list of classifier results --
+  # where a classifier result is a list of strings, 
+  # one string (= assigned class) for each instance,
+  # and where each list of classifier results has the same length --
+  # for each instance, combine individual classifier results
+  # into a single judgement
+  #
+  # returns: an array of strings: one combined classifier result, 
+  # one string (=assigned class) for each instance
+  def combine(classifier_results) #array:array:string, list of classifier results
+
+    if classifier_results.length() == 1
+      return classifier_results.first
+    elsif classifier_results.length() == 0
+      raise "Can't do classification with zero classifiers."
+    else
+      raise "True classifier combination not implemented yet"
+    end
+  end
+end
+
+
+##########################################################################
+# main class in this package:
+# applying classifiers
+class RosyTest < RosyTask
+
+  #####
+  # new:
+  # 
+  # initialize everything for applying classifiers
+  #
+  # argrec_apply: apply trained argrec classifiers to
+  # training data, which means that almost everything is different
+  def initialize(exp,      # RosyConfigData object: experiment description
+		 opts,     # hash: runtime argument option (string) -> value (string)
+		 ttt_obj,  # RosyTrainingTestTable object
+                 argrec_apply = false) # boolean. true: see above
+    
+    ##
+    # remember the experiment description
+
+    @exp = exp
+    @ttt_obj = ttt_obj
+    @argrec_apply = argrec_apply
+
+    ##
+    # check runtime options
+
+    # defaults:
+    @step = "both"
+    @splitID = nil
+    @testID = default_test_ID()
+    @produce_output = true
+
+    opts.each { |opt,arg|
+      case opt
+      when "--step"
+	unless ["argrec", "arglab", "both", "onestep"].include? arg
+	  raise "Classification step must be one of: argrec, arglab, both, onestep. I got: " + arg.to_s
+	end
+	@step = arg
+
+      when "--logID"
+        @splitID = arg
+
+      when "--testID"
+        @testID = arg
+
+      when "--nooutput"
+        @produce_output = false
+
+      else
+	# this is an option that is okay but has already been read and used by rosy.rb
+      end	
+    }
+
+    ##
+    # check: if this is about a split, do we have it?
+    # if it is about a test, do we have it?
+    if @splitID
+      unless @ttt_obj.splitIDs().include?(@splitID)
+        $stderr.puts "Sorry, I have no data for split ID #{@splitID}."
+        exit 1
+      end
+    else
+      if not(@argrec_apply) and not(@ttt_obj.testIDs().include?(@testID))
+        $stderr.puts "Sorry, I have no data for test ID #{@testID}."
+        exit 1
+      end
+    end
+
+    ##
+    # determine classifiers
+    #
+    # get_lf returns: array of pairs [classifier_name, options[array]]
+    #
+    # @classifiers: list of pairs [Classifier object, classifier name(string)]
+    @classifiers = @exp.get_lf("classifier").map { |classif_name, options|
+      [Classifier.new(classif_name, options), classif_name]
+    }
+    # sanity check: we need at least one classifier
+    if @classifiers.empty?
+      raise "I need at least one classifier, please specify using exp. file option 'classifier'"
+    end
+    
+    # make classifier combination object
+    @combinator = ClassifierCombination.new(@exp)
+
+    if not(@argrec_apply)
+      # normal run
+
+      #####
+      # Enduser mode: only steps "both" and "onestep" available.
+      # testing only on test data, not on split data
+      in_enduser_mode_ensure(["both", "onestep"].include?(@step))
+
+      ##
+      # add preprocessing information to the experiment file object
+      # @note AB: Commented out due to separation of PrepConfigData:
+      #   information for SynInteraces required.
+      # if @splitID
+      #   # use split data
+      #   preproc_param = "preproc_descr_file_train"
+      # else
+      #   # use test data
+      #   preproc_param = "preproc_descr_file_test"
+      # end
+
+      # preproc_expname = @exp.get(preproc_param)
+      # if not(preproc_expname)
+      #   $stderr.puts "Please set the name of the preprocessing exp. file name"
+      #   $stderr.puts "in the experiment file, parameter #{preproc_param}."
+      #   exit 1
+      # elsif not(File.readable?(preproc_expname))
+      #   $stderr.puts "Error in the experiment file:"
+      #   $stderr.puts "Parameter #{preproc_param} has to be a readable file."
+      #   exit 1
+      # end
+      # preproc_exp = FrPrepConfigData.new(preproc_expname)
+      # @exp.adjoin(preproc_exp)
+
+      # announce the task
+      $stderr.puts "---------"
+      $stderr.print "Rosy experiment #{@exp.get("experiment_ID")}: Testing "
+      if @splitID
+        $stderr.puts "on split dataset #{@splitID}"
+      else
+        $stderr.puts "on test dataset #{@testID}"
+      end
+      $stderr.puts "---------"
+    end
+  end
+
+
+  ##################################################################
+  # perform
+  #
+  # apply trained classifiers to the given (test) data
+  def perform()
+    if @step == "both"
+      # both? then do first argrec, then arglab
+      $stderr.puts "Rosy testing step argrec"
+      
+      previous_produce_output = @produce_output # no output in argrec
+      @produce_output = false  # when performing both steps in a row
+
+      @step = "argrec"
+      perform_aux()
+
+      $stderr.puts "Rosy testing step arglab"
+      @produce_output = previous_produce_output
+      @step = "arglab"
+      perform_aux()
+    else
+      # not both? then just do one
+      $stderr.puts "Rosy testing step " + @step
+      perform_aux()
+    end
+
+    ####
+    # Enduser mode: remove DB table with test data
+    if $ENDUSER_MODE
+      $stderr.puts "---"
+      $stderr.puts "Cleanup: Removing DB table with test data."
+
+      unless @testID
+        raise "Shouldn't be here"
+      end
+
+      @ttt_obj.remove_test_table(@testID)
+    end
+  end
+
+  ######################
+  # get_result_column_name
+  #
+  # returns the column name for the current run,
+  # i.e. the name of the column where this object's perform method
+  # writes its data
+  def get_result_column_name()
+    return @run_column
+  end
+
+  #################################
+  private
+
+  # perform_aux: do the actual work of the perform() method
+  # moved here because of the possibility of having @step=="both",
+  # which makes it necessary to perform two test steps one after the other
+  def perform_aux()
+
+    @iterator, @run_column = get_iterator(true)
+
+    ####
+    # get the list of relevant features,
+    # remove the features that describe the unit by which we train, 
+    # since they are going to be constant throughout the training file
+    
+    @features = @ttt_obj.feature_info.get_model_features(@step) - 
+                @iterator.get_xwise_column_names()
+
+    # but add the gold feature
+    unless @features.include? "gold"
+      @features << "gold"
+    end
+    
+    ####
+    # for each group (as defined by the @iterator):
+    # apply the group-specific classifier,
+    # write the result into the database, into 
+    # the column named @run_column
+    classif_dir = classifier_directory_name(@exp, @step, @splitID)
+
+    @iterator.each_group { |group_descr_hash, group|
+
+      $stderr.puts "Applying classifiers to: " + group.to_s
+      
+      # get data for current group from database:
+      
+      # make a view: model features
+      feature_view = @iterator.get_a_view_for_current_group(@features)
+      
+	if feature_view.length() == 0
+        # no test data in this view: next group
+        feature_view.close()
+        next
+      end
+      
+      # another view for writing the result
+      result_view = @iterator.get_a_view_for_current_group([@run_column])
+
+      # read trained classifiers
+      # classifiers_read_okay: boolean, true if reading the stored classifier(s) succeeded
+      classifiers_read_okay = true
+      
+      @classifiers.each { |classifier, classifier_name| 
+        
+        stored_classifier = classif_dir + 
+              @exp.instantiate("classifier_file",
+                               "classif" => classifier_name,
+                                       "group" => group.gsub(/ /, "_"))
+        
+        status = classifier.read(stored_classifier)
+        unless status
+          STDERR.puts "[RosyTest] Error: could not read classifier."
+          classifiers_read_okay = false
+        end
+        
+      }
+
+      classification_result = Array.new
+      
+      if classifiers_read_okay        
+        # apply classifiers, write result to database
+        classification_result = apply_classifiers(feature_view, group, "test")
+      end
+      
+      if classification_result == Array.new
+        # either classifiers did not read OK, or some problem during classification: 
+        # label everything with NONE
+        result_view.each_instance_s {|inst|
+          classification_result << @exp.get("noval")
+        }        
+      end
+
+      result_view.update_column(@run_column, 
+                                classification_result)
+      feature_view.close()
+      result_view.close()
+    }
+
+    # pruning? then set the result for pruned nodes to "noval"
+    # if we are doing argrec or onestep
+    integrate_pruning_into_argrec_result()
+
+    # postprocessing:
+    # remove superfluous role labels, i.e. labels on nodes
+    # whose ancestors already bear the same label
+    if @step == "argrec" or @step == "onestep"
+
+      $stderr.puts "Postprocessing..."
+
+      # iterator for doing the postprocessing:
+      # no pruning
+      @postprocessing_iterator, dummy = get_iterator(false)
+
+      @postprocessing_iterator.each_group { |group_descr_hash, group|
+
+	view = @postprocessing_iterator.get_a_view_for_current_group(["nodeID", "sentid", @run_column])
+	
+	# remove superfluous labels, write the result back to the DB
+	postprocess_classification(view, @run_column)
+	view.close()
+      }
+    end
+
+
+    # all went well, so confirm this run
+    if @argrec_apply
+      # argrec_apply: don't add preprocessing info again, and 
+      # get view maker for the training data
+      @ttt_obj.confirm_runlog("argrec", "train", @testID, @splitID, @run_column)
+    else
+      # normal run
+      @ttt_obj.confirm_runlog(@step, "test", @testID, @splitID, @run_column)
+    end
+
+    ####
+    # If we are being asked to produce SalsaTigerXML output:
+    # produce it.
+    if @produce_output
+      write_stxml_output()
+    end
+  end
+
+  #########################
+  # returns a pair [iterator, run_column]
+  # for the current settings
+  #
+  # prune = true: If pruning has been enabled,
+  # RosyIterator will add the appropriate DB column restrictions
+  # such that pruned constituents do nto enter into training
+  def get_iterator(prune)  #Boolean
+    ##
+    # make appropriate iterator object, get column name for the current run
+    #
+    if @argrec_apply
+      # get view maker for the training data
+      iterator = RosyIterator.new(@ttt_obj, @exp, "train", 
+                                   "step" => @step, 
+                                   "splitID" => @splitID,
+                                   "prune" => prune)
+      run_column = @ttt_obj.new_runlog("argrec", "train", @testID, @splitID)
+
+    else
+      # normal run
+
+      # hand all the info to the RosyIterator object
+      # It will figure out what view I'll need
+      iterator = RosyIterator.new(@ttt_obj, @exp, "test", 
+                                  "step" => @step, 
+                                  "testID" => @testID, 
+                                  "splitID" => @splitID,
+                                  "prune" => prune)
+
+      run_column = @ttt_obj.new_runlog(@step, "test", @testID, @splitID)
+    end
+
+    return [iterator, run_column]
+  end
+
+  #########################
+  # integrate pruning result into argrec result
+  def integrate_pruning_into_argrec_result()
+    if ["argrec", "onestep"].include? @step
+      # we only need to integrate pruning results into argument recognition
+
+      # get iterator that doesn't do pruning
+      iterator, run_column = get_iterator(false)
+      Pruning.integrate_pruning_into_run(run_column, iterator, @exp)
+    end
+  end
+
+  #########################
+  def apply_classifiers(view,  # DBView object: data to be classified
+                        group,       # string: frame or target POS we are classifying
+                        dataset)     # string: train/test
+                        
+    # make input file for classifiers
+    tf_input = Tempfile.new("rosy")
+    view.each_instance_s { |instance_string|
+      # change punctuation to _PUNCT_
+      # and change empty space to _
+      # because otherwise some classifiers may spit
+      tf_input.puts prepare_output_for_classifiers(instance_string)
+    }
+    tf_input.close()
+    # make output file for classifiers
+    tf_output = Tempfile.new("rosy")
+    tf_output.close()
+    
+    ###
+    # apply classifiers
+    
+    # classifier_results: array:array of strings, a list of classifier results,
+    # each result a list of assigned classes(string), one class for each instance of the view
+    classifier_results = Array.new
+
+    @classifiers.each { |classifier, classifier_name|
+
+
+      # did we manage to classify the test data?      
+      # there may be errors on the way (eg no training data)      
+      
+      success = classifier.apply(tf_input.path(), tf_output.path())
+      
+      if success
+        
+        # read classifier output from file
+        classifier_results << classifier.read_resultfile(tf_output.path()).map { |instance_result|
+          # instance_result is a list of pairs [label, confidence]
+          # such that the label with the highest confidence is first
+          if instance_result.empty?
+            # oops, no results
+            nil
+          else
+            # label of the first label/confidence pair
+            instance_result.first().first()
+          end
+        }.compact()
+        
+      else
+        # error: return empty Array, so that error handling can take over in perform_aux()
+        return Array.new
+      end
+    }
+
+    # if we are here, all classifiers have succeeded... 
+    
+    # clean up
+    tf_input.close(true)
+    tf_output.close(true)
+
+    # combine classifiers
+    return @combinator.combine(classifier_results)
+  end
+
+  ###
+  # postprocess_classification
+  #
+  # given output of a learner,
+  # postprocess the output:
+  # map cases of
+  #    FE
+  #   /  \
+  #       ...
+  #       \
+  #        FE
+  #
+  # to 
+  #    FE
+  #   /  \
+  #       ...
+  #        \
+  #        NONE
+  def postprocess_classification(view, # DBView object: node IDs
+                                 run_column) # string: name of current run column
+
+
+    # keep new values for run_column for all rows in view
+    # will be used for update in the end
+    result = Array.new()
+
+    view.each_sentence() { |sentence|
+
+      # returns hash: 
+      # node index -> array of node indices: ancestors of the given node
+      # indices are indices in the 'sentence' array
+      ancestors = make_ancestor_hash(sentence)
+
+      # test output
+#       $stderr.puts "nodeID values:"
+#       sentence.each_with_index  { |inst, index| 
+#         $stderr.puts "#{index}) #{inst["nodeID"]}"
+#       }
+#       $stderr.puts "\nAncestor hash:"
+#       ancestors.each_pair { |node_ix, ancestors|
+#         $stderr.puts "#{node_ix} -> " +  ancestors.map { |a| a.to_s }.join(", ")
+#       }
+#       $stderr.puts "press enter"
+#       $stdin.gets()
+
+      sentence.each_with_index { |instance, inst_index|
+
+	# check whether this instance has an equally labeled ancestor
+	has_equally_labeled_ancestor = false
+
+	if (instance[run_column] != @exp.get("noval")) and 
+	  ancestors[inst_index]
+	  
+	  if ancestors[inst_index].detect { |anc_index|
+	      sentence[anc_index][run_column] == instance[run_column]
+	    }
+	    has_equally_labeled_ancestor = true
+	  else
+	    has_equally_labeled_ancestor = false
+	  end
+	end
+
+
+	if has_equally_labeled_ancestor
+	  result << @exp.get("noval")
+	else
+	  result << instance[run_column]
+	end
+      }
+    }
+
+
+#     # checking: how many labels have we deleted?
+#     before = 0
+#     view.each_sentence { |s|
+#       s.each { |inst| 
+# 	unless inst[run_column] == @exp.get("noval")
+# 	  before += 1
+# 	end
+#       }
+#     }
+#     after = 0
+#     result.each { |r| 
+#       unless r == @exp.get("noval")
+# 	after += 1
+#       end
+#     }
+#     $stderr.puts "Non-NONE labels before: #{before}"
+#     $stderr.puts "Non-NONE labels after: #{after}"
+
+
+    # update DB to new result
+    view.update_column(run_column, result)
+  end
+
+  ##
+  # make_ancestor_hash
+  #
+  # given a sentence as returned by view.each_sentence
+  # (an array of hashes: column_name -> column_value),
+  # use the column nodeID to map each instance of the sentence to its
+  # ancestors
+  #
+  # returns: hash instanceID(integer) -> array:instanceIDs(integers)
+  # mapping each instance to the list of its ancestors
+  def make_ancestor_hash(sentence) # array:hash: column_name(string) -> column_value(object)
+    # for each instance: find the parent
+    # and store it in the parent_index hash
+    parent_index = Hash.new
+
+
+    # first make hash mapping each node ID to its index in the 
+    # 'sentence' array
+    id_to_index = Hash.new()
+    sentence.each_with_index { |instance, index|
+      if instance["nodeID"]
+        myID, parentID = instance["nodeID"].split()
+        id_to_index[myID] = index
+      else
+        $stderr.puts "WARNING: no node ID for instance:\n"
+        $stderr.puts instance.values.join(",")
+      end
+    }
+
+    # now make hash mapping each node index to its parent index
+    sentence.each { |instance|
+      if instance["nodeID"]
+        myID, parentID = instance["nodeID"].split()
+        if parentID # root has no parent ID
+
+          # sanity check: do I know the indices?
+          if id_to_index[myID] and id_to_index[parentID]
+            parent_index[id_to_index[myID]] = id_to_index[parentID]
+          else
+            $stderr.puts "RosyTest postprocessing WARNING: found ID for unseen nodes"
+          end
+        end
+      else
+        $stderr.puts "RosyTest postprocessing WARNING: no node ID for instance:\n"
+        $stderr.puts instance.values.join(",")
+      end
+    }
+
+    # for each instance: gather ancestor IDs
+    # and store them in the ancestor_index hash
+    ancestor_index = Hash.new
+
+    parent_index.each_key { |node_index|
+      ancestor_index[node_index] = Array.new
+      ancestor = parent_index[node_index]
+
+      while ancestor
+        if ancestor_index[node_index].include? ancestor 
+          # we seem to have run into a loop
+          # this should not happen, but it has happened anyway ;-)
+#          STDERR.puts "Warning: node #{ancestor} is its own ancestor!"
+          break
+        end
+        ancestor_index[node_index] << ancestor
+        ancestor = parent_index[ancestor]
+      end
+    }
+    return ancestor_index
+  end
+
+  ################
+  # write_stxml_output
+  #
+  # Output the result of Rosy as SalsaTigerXML:
+  # Take the input SalsaTigerXML data,
+  # and write them to directory_output 
+  # (or, lacking that, to <rosy_dir>/<experiment_ID>/output),
+  # taking over the frames from the input data
+  # and supplanting any FEs that might be set in the input data
+  # by the ones newly assigned by Rosy.
+  def write_stxml_output()
+
+    ##
+    # determine input and output directory
+    rosy_dir = File.new_dir(@exp.instantiate("rosy_dir",
+                                             "exp_ID" => @exp.get("experiment_ID")))
+    if @splitID
+      # split data is being used: part of the training data
+      input_directory = File.existing_dir(rosy_dir,"input_dir/train")
+    else
+      # test data is being used
+      input_directory = File.existing_dir(rosy_dir, "input_dir/test")
+    end
+
+    
+    if @exp.get("directory_output")
+      # user has set an explicit output directory
+      output_directory = File.new_dir(@exp.get("directory_output"))
+    else
+      # no output directory has been set: use default
+      output_directory = File.new_dir(@exp.instantiate("rosy_dir", "exp_ID" => @exp.get("experiment_ID")),
+                                      "output")
+    end    
+    
+    ###
+    # find appropriate class for interpreting syntactic structures
+    interpreter_class = SynInterfaces.get_interpreter_according_to_exp(@exp)
+
+
+    $stderr.puts "Writing SalsaTigerXML output to #{output_directory}"
+
+    ###
+    # read in all FEs that have been assigned
+    # sentid_to_assigned: hash <sent ID, frame ID> (string) -> array of pairs [FE, node ID]
+    sentid_to_assigned = Hash.new
+    @iterator.each_group { |group_descr_hash, group|
+      view = @iterator.get_a_view_for_current_group([@run_column, "nodeID", "sentid"])
+
+      view.each_hash { |inst_hash|
+        # if this sentence ID/frame ID pair is in the test data,
+        # its hash entry will at least be nonnil, even if no 
+        # FEs have been assigned for it
+        unless sentid_to_assigned[inst_hash["sentid"]]
+          sentid_to_assigned[inst_hash["sentid"]] = Array.new
+        end
+
+        # if nothing has been assigned to this instance, don't record it
+        if inst_hash[@run_column].nil? or inst_hash[@run_column] == @exp.get("noval")
+          next
+        end
+
+        # record instance
+        sentid_to_assigned[inst_hash["sentid"]] << [inst_hash[@run_column], inst_hash["nodeID"]]
+      }
+      view.close()
+    }
+
+    ###
+    # write stuff
+
+    ##
+    # iterate through input files
+    Dir[input_directory + "*.xml.gz"].each { |infilename|
+
+      # unpack input file
+      tempfile = Tempfile.new("RosyTest")
+      tempfile.close()
+      %x{gunzip -c #{infilename} > #{tempfile.path()}}
+
+      # open input and output file
+      infile = FilePartsParser.new(tempfile.path())
+      outfilename = output_directory + File.basename(infilename, ".gz")
+      begin
+        outfile = File.new(outfilename, "w")
+      rescue
+        raise "Could not write to SalsaTigerXML output file #{outfilename}"
+      end
+
+      # write header to output file
+      outfile.puts infile.head()
+      
+      ##
+      # each input sentence: integrate newly assigned roles
+      infile.scan_s { |sent_string|
+        sent = SalsaTigerSentence.new(sent_string)
+        
+        ##
+        # each input frame: remove old roles, add new ones
+        sent.frames.each { |frame|
+
+          # this corresponds to the sentid feature in the database
+          sent_frame_id = construct_instance_id(sent.id(), frame.id())
+
+          if sentid_to_assigned[sent_frame_id].nil? and @splitID
+	    # we are using a split of the training data, and
+            # this sentence/frame ID pair does not
+            # seem to be in the test part of the split
+            # so do not show the frame
+	    #
+	    # Note that if we are _not_ working on a split,
+	    # we are not discarding any frames or sentences
+            sent.remove_frame(frame)
+          end
+
+          # remove old roles, but do not remove target
+          old_fes = frame.children()
+          old_fes.each { |old_fe|
+            unless old_fe.name() == "target"
+              frame.remove_child(old_fe)
+            end
+          }
+
+          if sentid_to_assigned[sent_frame_id].nil?
+            # nothing assigned to this frame -- go on
+            next
+          end
+
+          # assign new roles:
+          # each FE occurring for this sentence ID plus frame ID:
+          # collect all node ID / parentnode ID pairs listed for that FE,
+          # map the IDs to actual nodes, and assign the FE.
+          sentid_to_assigned[sent_frame_id].map { |fe_name, npp| fe_name }.uniq.each { |fe_name|
+            # each FE
+
+            nodes = sentid_to_assigned[sent_frame_id].select { |other_fe_name, npp| 
+              # collect node ID / parentnode ID pairs listed for that FE
+              other_fe_name == fe_name
+
+            }.map { |other_fe_name, nodeid_plus_parent_id|
+              # map the node ID / parentnode ID pair to an actual node
+
+              node_id, parent_id = nodeid_plus_parent_id.split()
+              if node_id == @exp.get("noval")
+                $stderr.puts "Warning: got NONE for a node ID"
+                node = nil
+
+              else
+                node = sent.syn_node_with_id(node_id)
+                unless node
+                  $stderr.puts "Warning: could not find node with ID #{node_id}"
+                end
+              end
+
+              node
+            }.compact
+            
+            # assign the FE
+            sent.add_fe(frame, fe_name, interpreter_class.max_constituents(nodes, sent))
+          } # each FE
+        } # each frame
+
+        # write changed sentence to output file
+        # if we are working on a split of the training data,
+	# write the sentence only if there are frames in it
+	if sent.frames.length() == 0 and @splitID
+	  # split of the training data, and no frames
+	else
+          outfile.puts sent.get()
+        end
+      } # each sentence
+
+      # write footer to output file
+      outfile.puts infile.tail()
+      tempfile.close(true)
+    } # each input file
+  end
+end