shalmaneser 0.0.1.alpha

data/lib/fred/FredFeatures.rb

@@ -0,0 +1,1061 @@
+# -*- coding: utf-8 -*-
+require "tempfile"
+require "delegate"
+
+require "fred/FredFeatureExtractors"
+
+########################################
+########################################
+# Feature access classes:
+# read and write features
+class AbstractFredFeatureAccess
+  ####
+  def initialize(exp, # experiment file object
+                 dataset, # dataset: "train" or "test"
+                 mode = "r") # mode: r, w, a
+    @exp = exp
+    @dataset = dataset
+    @mode = mode
+    
+    unless ["r", "w", "a"].include? @mode
+      $stderr.puts "FeatureAccess: unknown mode #{@mode}."
+      exit 1
+    end
+
+  end
+
+  ####
+  def AbstractFredFeatureAccess.remove_feature_files()
+    raise "overwrite me"
+  end
+
+  ####
+  def write_item(lemma,  # string: lemma
+                 pos,    # string: POS
+                 ids,    # array:string: unique IDs of this occurrence of the lemma
+                 sid,    # string: sentence ID
+                 sense,  # string: sense
+                 features) # features: hash feature type-> features (string-> array:string)
+    raise "overwrite me"
+  end
+                 
+
+  def flush()
+    raise "overwrite me"
+  end
+end
+
+########################################
+# MetaFeatureAccess:
+# write all featurization data to one gzipped file,
+# directly writing the meta-features as they come
+# format:
+#
+# lemma pos id sense
+#   <feature_type>: <features>
+#
+# where feature_type is a word, and features is a list of words, space-separated
+class MetaFeatureAccess < AbstractFredFeatureAccess
+  ###
+  def initialize(exp, dataset, mode)
+    super(exp, dataset, mode)
+
+    @filename = MetaFeatureAccess.filename(@exp, @dataset)
+
+    # make filename for writing features
+    case @mode
+
+    when "w", "a", "r"
+      # read or write access
+      @f = FileZipped.new(@filename, mode)
+
+    else
+      $stderr.puts "MetaFeatureAccess error: illegal mode #{mode}"
+      exit 1
+    end
+  end
+
+
+  ####
+  def MetaFeatureAccess.filename(exp, dataset, mode="new")
+    return fred_dirname(exp, dataset, "meta_features", mode) +
+      "meta_features.txt.gz"
+  end
+
+  ####
+  def MetaFeatureAccess.remove_feature_files(exp, dataset)
+    filename = MetaFeatureAccess.filename(exp, dataset)
+    if File.exists?(filename)
+      File.delete(filename)
+    end
+  end
+
+
+  ###
+  # read items, yield one at a time
+  #
+  # format: tuple consisting of
+  # - target_lemma: string
+  # - target_pos: string
+  # - target_ids: array:string
+  # - target SID: string, sentence ID
+  # - target_senses: array:string
+  # - feature_hash: feature_type->values, string->array:string
+  def each_item()
+    unless @mode == "r"
+      $stderr.puts "MetaFeatureAccess error: cannot read file not opened for reading"
+      exit 1
+    end
+
+    lemma = pos = sid = ids = senses = nil
+
+    feature_hash = Hash.new()
+
+    @f.each { |line|
+      line.chomp!
+      if line =~ /^\s/
+        # line starts with whitespace: continues description of previous item
+        # that is, if we have a previous item
+        #
+        # format of line:
+        #    feature_type: feature feature feature ...
+        # as in 
+        #    CH: SB#expansion#expansion#NN# OA#change#change#NN#
+        unless lemma
+          $stderr.puts "MetaFeatureAccess error: unexpected leading whitespace"
+          $stderr.puts "in meta-feature file #{@filename}, ignoring line:"
+          $stderr.puts line
+          next
+        end
+
+        feature_type, *features = line.split()
+        
+        unless feature_type =~ /^(.*):$/
+          # feature type should end in ":"
+          $stderr.puts "MetaFeatureAccess error: feature type should end in ':' but doesn't"
+          $stderr.puts "in meta-feature file #{@filename}, ignoring line:"
+          $stderr.puts line
+          next
+        end
+
+        feature_hash[feature_type[0..-2]] = features
+         
+        
+      else
+        # first line of item.
+        #
+        # format:
+        # lemma POS IDs sid senses
+        # 
+        # as in:
+        # cause verb 2-651966_8 2-651966 Causation
+
+        # first yield previous item
+        if lemma
+          yield [lemma, pos, ids, sid, senses, feature_hash]
+        end
+
+        # then start new item:
+        lemma, pos, ids_s, sid, senses_s = line.split()
+        ids = ids_s.split("::").map { |i| i.gsub(/COLON/, ":") }
+        senses = senses_s.split("::").map { |s| s.gsub(/COLON/, ":") }
+
+        # reset feature hash
+        feature_hash.clear()
+      end
+    }
+
+    # one more item to yield?
+    if lemma
+      yield [lemma, pos, ids, sid, senses, feature_hash]
+    end
+  end
+
+
+
+  ###
+  def write_item(lemma,  # string: target lemma
+                 pos,    # string: target pos
+                 ids,    # array:string: unique IDs of this occurrence of the lemma
+                 sid,    # string: sentence ID
+                 senses, # array:string: sense
+                 features) # features: hash feature type-> features (string-> array:string)
+
+    unless ["w", "a"].include? @mode
+      $stderr.puts "MetaFeatureAccess error: cannot write to feature file opened for reading"
+      exit 1
+    end
+
+    if not(lemma) or lemma.empty? or not(ids) or ids.empty?
+      # nothing to write
+      # HIER debugging
+      # raise "HIER no lemma or no IDs: #{lemma} #{ids}"
+      return
+    end
+    if pos.nil? or pos.empty?
+      # POS unknown
+      pos = ""
+    end
+    unless senses
+      senses = [ @exp.get("noval") ]
+    end
+
+    ids_s = ids.map { |i| i.gsub(/:/, "COLON") }.join("::")
+
+    senses_s = senses.map { |s| s.gsub(/:/, "COLON") }.join("::")
+    @f.puts "#{lemma} #{pos} #{ids_s} #{sid} #{senses_s}"
+    features.each_pair { |feature_type, f_list|
+      @f.puts "   #{feature_type}: " + f_list.map { |f| f.to_s() }.join(" ")
+    }
+    @f.flush()
+  end
+
+  ###
+  def flush()
+    unless ["w", "a"].include? @mode
+      $stderr.puts "MetaFeatureAccess error: cannot write to feature file opened for reading"
+      exit 1
+    end
+
+    # actually, nothing to be done here
+  end
+
+end
+
+
+########################################
+# FredFeatureWriter:
+# write chosen features (according to the experiment file)
+# to
+# - one file per lemma for n-ary classification
+# - one file per lemma/sense pair for binary classification
+#
+# format: CSV, last entry is target class
+class FredFeatureAccess < AbstractFredFeatureAccess
+  ###
+  def initialize(exp, dataset, mode)
+    super(exp, dataset, mode)
+
+    # write to auxiliary files first,
+    # to sort items by lemma
+    @w_tmp = AuxKeepWriters.new()
+
+    # which features has the user requested?
+    feature_info_obj = FredFeatureInfo.new(@exp)
+    @feature_extractors = feature_info_obj.get_extractor_objects()
+
+  end
+
+  ####
+  def FredFeatureAccess.remove_feature_files(exp, dataset)
+    
+    # remove feature files
+    WriteFeaturesNaryOrBinary.remove_files(exp, dataset)
+
+    # remove key files
+    AnswerKeyAccess.remove_files(exp, dataset)
+  end
+  
+  ###
+  def  FredFeatureAccess.legend_filename(lemmapos)
+    return "fred.feature_legend.#{lemmapos}"
+  end
+
+  ### 
+  def FredFeatureAccess.feature_dir(exp, dataset)
+    return WriteFeaturesNaryOrBinary.feature_dir(exp, dataset, "new")
+  end
+
+  ###
+  # each feature file:
+  # iterate through feature files,
+  # yield pairs [filename, values]
+  # where 'values' is a hash containing keys
+  # 'lemma' and potentially 'sense'
+  #
+  # filenames are sorted alphabetically before being yielded
+  #
+  # available in read and write mode
+  def FredFeatureAccess.each_feature_file(exp, dataset)
+    feature_dir = FredFeatureAccess.feature_dir(exp, dataset)
+    Dir[feature_dir + "*"].sort().each { |filename|
+      if (values = deconstruct_fred_feature_filename(filename))
+        yield [filename, values]
+      end
+    }
+  end
+
+  ###
+  # write item: 
+  # - transform meta-features into actual features as requested
+  #   in the experiment file
+  # - write item to tempfile, don't really write yet
+  def write_item(lemma,  # string: target lemma
+                 pos,    # string: target pos
+                 ids,    # array:string: unique IDs of this occurrence of the lemma
+                 sid,    # string: sentence ID
+                 senses,  # array:string: sense
+                 features) # features: hash feature type-> features (string-> array:string)
+
+
+    unless ["w", "a"].include? @mode
+      $stderr.puts "FredFeatures error: cannot write to feature file opened for reading"
+      exit 1
+    end
+
+    if lemma.nil? or lemma.empty? or ids.nil? or ids.empty?
+      # nothing to write
+      return
+    end
+    if pos.nil? or pos.empty?
+      # POS unknown
+      pos = ""
+    end
+
+    # falsch! noval nicht zulässig für fred! (nur für rosy!) - Warum steht das hier???
+    unless senses
+      senses = [ @exp.get("noval") ]
+    end
+
+    # modified by ines, 19.7.2010
+    # senses should be empty, but they are not - why?
+    if senses.length == 1 and senses[0].eql? "" 
+	senses = "NONE" 
+    end
+
+    writer = @w_tmp.get_writer_for(fred_lemmapos_combine(lemma, pos))
+    ids_s = ids.map { |i| i.gsub(/:/, "COLON") }.join("::")
+    
+    # AB: Ines modified <senses> and it can be a String.
+    # That's corrected, but I do not guarantee the correct results.
+    if senses.respond_to? :map
+      senses_s = senses.map { |s| s.gsub(/:/, "COLON") }.join("::")
+    end
+    writer.print "#{lemma} #{pos} #{ids_s} #{sid} #{senses_s} "
+
+    # write all features
+    @feature_extractors.each { |extractor|
+      extractor.each_feature(features) { |feature|
+        writer.print feature, " "
+      }
+    }
+    writer.puts
+    writer.flush()
+  end
+
+  ###
+  def flush()
+    unless ["w", "a"].include? @mode
+      $stderr.puts "FredFeatureAccess error: cannot write to feature file opened for reading"
+      exit 1
+    end
+
+    # elements in the feature vector: get fixed with the training data, 
+    # get read with the test data.
+    # get stored in feature_legend_dir
+    case @dataset
+    when "train"
+      feature_legend_dir = File.new_dir(fred_classifier_directory(@exp), 
+                                        "legend")
+    when "test"
+      feature_legend_dir= File.existing_dir(fred_classifier_directory(@exp), 
+                                        "legend")
+    end      
+
+    # now really write features
+    @w_tmp.flush()
+    @w_tmp.get_lemmas().sort().each { |lemmapos|
+   
+      # inform user
+      $stderr.puts "Writing #{lemmapos}..."
+
+      # prepare list of features to use in the feature vector:
+      legend_filename = feature_legend_dir + FredFeatureAccess.legend_filename(lemmapos)
+
+      case @dataset
+      when "train"
+        # training data: 
+        # determine feature list and sense list from the data,
+        # and store in the relevant file
+        feature_list, sense_list = collect_feature_list(lemmapos)
+        begin
+          f = File.new(legend_filename, "w")
+        rescue
+          $stderr.puts "Error: Could not write to feature legend file #{legend_filename}: " + $!
+          exit 1
+        end
+        f.puts feature_list.map { |x| x.gsub(/,/, "COMMA") }.join(",")
+        f.puts sense_list.map { |x| x.gsub(/,/, "COMMA") }.join(",")
+        f.close()
+
+      when "test"
+        # test data:
+        # read feature list and sense list from the relevant file
+      
+        begin
+          f = File.new(legend_filename)
+        rescue
+          $stderr.puts "Error: Could not read feature legend file #{legend_filename}: " + $!
+          $stderr.puts "Skipping this lemma."
+          next
+        end
+        feature_list = f.gets().chomp().split(",").map { |x| x.gsub(/COMMA/, ",") }
+        sense_list = f.gets().chomp().split(",").map { |x| x.gsub(/COMMA/, ",") }
+      end
+
+      # write
+      # - featurization file
+      # - answer key file
+
+      f = @w_tmp.get_for_reading(lemmapos)
+      answer_obj = AnswerKeyAccess.new(@exp, @dataset, lemmapos, "w")
+
+      obj_out = WriteFeaturesNaryOrBinary.new(lemmapos, @exp, @dataset)
+
+      f.each { |line| 
+
+        lemma, pos, ids, sid, senses, features = parse_temp_itemline(line)
+        unless lemma 
+          # something went wrong in parsing the line
+          next
+        end
+        each_sensegroup(senses, sense_list) { |senses_for_item, original_senses|
+          # write answer key
+          answer_obj.write_line(lemma, pos, 
+                                ids, sid, original_senses, senses_for_item)
+
+          # write item: features, senses
+	  obj_out.write_instance(to_feature_list(features, feature_list),
+				 senses_for_item)
+        } # each sensegroup
+      } # each input line
+      obj_out.close()
+      answer_obj.close()
+      @w_tmp.discard(lemmapos)
+    } # each lemma
+
+
+  end
+
+  ##################
+  protected
+
+  ###
+  # read temp feature file for the given lemma/pos
+  # and determine the list of all features and the list of all senses,
+  # each sorted alphabetically
+  def collect_feature_list(lemmapos)
+    # read entries for this lemma
+    f = @w_tmp.get_for_reading(lemmapos)
+
+    # keep a record of all senses and features
+    # senses: binary.
+    # features: keep the max. number of times a given feature occurred
+    #         in an instance
+    all_senses = Hash.new()
+    all_features = Hash.new(0)
+    features_this_instance = Hash.new(0)
+    # record how often each feature occurred all in all
+    num_occ = Hash.new(0)
+    num_lines = 0
+
+    f.each { |line|
+      lemma, pos, id_string, sid, senses, features = parse_temp_itemline(line)
+
+      unless lemma
+        # something went wrong in parsing the line
+        # print out the file contents for reference, then leave
+        $stderr.puts "Could not read temporary feature file #{f.path()} for #{lemmapos}."
+        exit 1
+      end
+      num_lines += 1
+      senses.each { |s| all_senses[s] = true }
+      features_this_instance.clear()
+      features.each { |fea| 
+        features_this_instance[fea] += 1 
+        num_occ[fea] += 1
+      }
+      
+      features_this_instance.each_pair { |feature, value|
+        all_features[feature] = [ all_features[feature], features_this_instance[feature] ].max()
+      }
+    }
+
+    # HIER 
+    # if num_lines > 2
+    #  num_occ.each_pair { |feature, num_occ|
+    #    if num_occ < 2
+    #      all_features.delete(feature)
+    #    end
+    #  }
+    # end
+    
+
+
+    case @exp.get("numerical_features")
+    when "keep"
+      # leave numerical features as they are, or 
+      # don't do numerical features
+      return [ all_features.keys().sort(),
+               all_senses.keys().sort()
+             ]
+
+    when "repeat"
+      # repeat: turn numerical feature with max. value N
+      # into N binary features
+      feature_list = Array.new()
+      all_features.keys().sort().each { |feature|
+        all_features[feature].times() { |index| 
+          feature_list << feature + " #{index}/#{all_features[feature]}" 
+        }
+      }
+      return [ feature_list, 
+               all_senses.keys().sort()
+             ]
+
+    when "bin"
+      # make bins:
+      # number of bins = (max. number of occurrences of a feature per item) / 10
+      feature_list = Array.new()
+      all_features.keys().sort().each { |feature|
+        num_bins_this_feature = (all_features[feature].to_f() / 10.0).ceil().to_i()
+
+        num_bins_this_feature.times { |index| 
+          feature_list << feature  + " #{index}/#{num_bins_this_feature}"
+        }
+      }
+      return [ feature_list,
+               all_senses.keys().sort()
+             ]
+    else
+      raise "Shouldn't be here"
+    end
+  end
+
+
+  ###
+  # given a full sorted list of items and a partial list of items,
+  # match the partial list to the full list,
+  # that is, produce as many items as the full list has
+  # yielding 0 where the partial entry is not in the full list, 
+  # and > 0 otherwise
+  #
+  # Note that if partial contains items not in full,
+  # they will not occur on the feature list returned!
+  def to_feature_list(partial, full,
+                      handle_numerical_features = nil)
+
+    #print "FULL: ", full, "\n"
+    #print "PART: ", partial, "\n"
+    # count occurrences of each feature in the partial list
+    occ_hash = Hash.new(0)
+    partial.each { |p|
+      occ_hash[p] += 1
+    }
+
+    # what to do with our counts?
+    unless handle_numerical_features
+      # no pre-set value given when this function was called
+      handle_numerical_features = @exp.get("numerical_features")
+    end
+
+    case handle_numerical_features
+    when "keep"
+      # leave numerical features as numerical features
+      return full.map { |x|
+        occ_hash[x].to_s()
+      }
+
+    when "repeat"
+      # repeat each numerical feature up to a max. number of occurrences
+      return full.map { |feature_plus_count|
+        unless feature_plus_count =~ /^(.*) (\d+)\/(\d+)$/
+          $stderr.puts "Error: could not parse feature: #{feature_plus_count}, bailing out."
+          raise "Shouldn't be here."
+        end
+
+        feature = $1
+        current_count = $2.to_i()
+        max_num = $3.to_i()
+
+        if occ_hash[feature] > current_count
+          1
+        else
+          0
+        end
+      }    
+
+    when "bin"
+      # group numerical feature values into N bins.
+      # number of bins varies from feature to feature
+      # each bin contains 10 different counts
+      return full.map { |feature_plus_count|
+        unless feature_plus_count =~ /^(.*) (\d+)\/(\d+)$/
+          $stderr.puts "Error: could not parse feature: #{feature_plus_count}, bailing out."
+          raise "Shouldn't be here."
+        end
+
+        feature = $1
+        current_count = $2.to_i()
+        max_num = $3.to_i()
+
+        if occ_hash[feature] % 10 > (10 * current_count)
+          1
+        else
+          0
+        end
+      }
+    else
+      raise "Shouldn't be here"
+    end
+  end
+
+
+  ###
+  # how to treat instances with multiple senses?
+  # - either write one item per sense
+  # - or combine all senses into one string
+  # - or keep as separate senses
+  #
+  # according to 'handle_multilabel' in the experiment file
+  #
+  # yields pairs of [senses, original_senses]
+  # both are arrays of strings
+  def each_sensegroup(senses, full_sense_list)
+    case @exp.get("handle_multilabel")
+    when "keep"
+      yield [senses, senses]
+    when "join"
+      yield [ [fred_join_senses(senses)], senses]
+    when "repeat"
+      senses.each { |s|
+        yield [ [s], senses]
+      }
+    when "binarize"
+      yield [ senses, senses ]
+    else
+      $stderr.puts "Error: unknown setting #{exp.get("handle_multilabel")}"
+      $stderr.puts "for 'handle_multilabel' in the experiment file."
+      $stderr.puts "Please choose one of 'binary', 'keep', 'join', 'repeat'"
+      $stderr.puts "or leave unset -- default is 'binary'."
+      exit 1
+    end
+  end
+
+  ###
+  def parse_temp_itemline(line)
+    lemma, pos, ids_s, sid, senses_s, *features = line.split()
+    # fix me! senses is empty, takes context features instead
+    unless senses_s
+      # features may be empty, but we need senses
+      $stderr.puts "FredFeatures Error in word sense item line: too short."
+      $stderr.puts ">>#{line}<<"
+      return nil
+    end
+    
+    ids = ids_s.split("::").map { |i| i.gsub(/COLON/, ":") }
+    senses = senses_s.split("::").map { |s| s.gsub(/COLON/, ":") }
+
+    return [lemma, pos, ids, sid, senses, features]
+  end
+
+end
+
+########################################
+# read and write access to answer key files
+# manages a single answer key file for a given lemma/POS pair
+class AnswerKeyAccess
+  ###
+  def initialize(exp,      # experiment file object
+                 dataset,  # "train", "test"
+                 lemmapos, # lemma + POS (one string)
+                 mode,     # "r", "w", "a"
+                 split_id = nil,
+                 split_dataset = nil)
+    unless ["r", "w", "a"].include? mode
+      $stderr.puts "FredFeatures error: AnswerKeyAccess initialized with mode #{mode}."
+      exit 1
+    end
+
+    @mode = mode
+    
+    answer_filename = fred_dirname(exp, dataset, "keys", "new") +
+      fred_answerkey_filename(lemmapos)
+
+    # are we reading the whole answer key file, or only the test part
+    # of a split of it?
+    if split_id
+      # we are accessing part of a split
+      # we can only do that when reading!
+      unless @mode == "r"
+        $stderr.puts "AnswerKeyAccess error: cannot access split answer file in write mode."
+        exit 1
+      end
+
+      # apply_split returns a closed temporary file
+      split_obj = FredSplitPkg.new(exp)
+      @f = split_obj.apply_split(answer_filename, lemmapos, split_dataset, split_id)
+      if @f.nil?
+        # the split_dataset part of the split doesn't contain any data
+        $stderr.puts "Warning: no #{split_dataset} data for lemma #{lemmapos}"
+      else
+        @f.open()
+      end
+
+    else
+      # we are reading the whole thing
+      begin
+        @f = File.new(answer_filename, @mode)
+      rescue
+        @f = nil
+      end
+    end
+  end
+
+  ###
+  def write_line(lemma,     # string: lemma
+                 pos,       # string: POS
+                 ids,       # array:string: target IDs
+                 sid,       # string: sentence ID
+                 senses,    # array:string: senses
+                 senses_this_item) # array:string: senses for this item
+    unless ["w", "a"].include? @mode
+      $stderr.puts "FredFeatures error: AnswerKeyAccess: cannot write in read mode."
+      exit 1
+    end
+    unless @f
+      raise "Shouldn't be here"
+    end
+    
+    # write answer key:
+    # lemma POS ID senses
+    if senses.include? nil or senses.include? ""
+      raise "empty sense"
+    end
+    if senses_this_item.include? nil or senses_this_item.include? ""
+      raise "empty sense for this item"
+    end
+
+    senses_s = senses.map { |s| s.gsub(/,/, "COMMA")}.join(",")
+    senses_ti_s = senses_this_item.map { |s| 
+      s.gsub(/,/, "COMMA")}.join(",")
+    id_s = ids.map { |i| i.gsub(/:/, "COLON") }.join("::")
+
+    @f.puts "#{lemma} #{pos} #{id_s} #{sid} #{senses_s} #{senses_ti_s}"
+  end
+
+  ###
+  # yield one line at a time:
+  # tuple (lemma, POS, ids, sentence_ID, all_assigned_senses, transformed_senses_for_this_item)
+  def each()
+    unless @mode == "r"
+      $stderr.puts "FredFeatures error: AnsewrKeyAccess: cannot read in write mode"
+    end
+    unless @f
+      # something went wrong during initialization:
+      # split didn't contain data
+      return
+    end
+
+    @f.each { |line|
+
+      lemma, pos, id_s, sid, senses_s, senses_this_item_s = line.split()
+      ids = id_s.split("::").map { |i| i.gsub(/COLON/, ":") }
+      senses = senses_s.split(",").map { |s| s.gsub(/COMMA/, ",") }
+
+      senses_this_item = senses_this_item_s.split(",").map { |s|
+        s.gsub(/COMMA/, ",") }
+
+      yield [lemma, pos, ids, sid, senses, senses_this_item]
+    }
+  end
+
+  ###
+  def close()
+    @f.close()
+  end
+
+  ###
+  def AnswerKeyAccess.remove_files(exp, dataset)
+    Dir[fred_dirname(exp, dataset, "keys", "new") + fred_answerkey_filename("*")].each { |filename|
+      if File.exists?(filename)
+        File.delete(filename)
+      end
+    }
+  end
+end
+
+
+####################3
+# keep writers: auxiliary class for FredFeatureAccess:
+# write to several files at a time
+# in tempfiles
+class AuxKeepWriters
+  def initialize()
+    @lemma2temp = Hash.new()
+    @size = 50
+    @writers = Array.new()
+  end
+
+
+  ##
+  def flush()
+    @writers.each { |lemmapos, writer|
+      writer.close()
+    }
+  end
+
+  ##
+  def get_lemmas()
+    return @lemma2temp.keys()
+  end
+
+  ##
+  def get_for_reading(lemmapos)
+    if @lemma2temp[lemmapos]
+      # we have a writer for this
+
+      @lemma2temp[lemmapos].close()
+      @lemma2temp[lemmapos].open()
+      return @lemma2temp[lemmapos]
+
+    else
+      # no writer for this
+      return nil
+    end
+  end
+  
+  ##
+  # finally close temp file, remove information for lemma/pos
+  def discard(lemmapos)
+    if @lemma2temp[lemmapos]
+      @lemma2temp[lemmapos].close(true)
+      @lemma2temp.delete(lemmapos)
+    end
+  end
+
+  ##
+  def get_writer_for(lemmapos)
+
+    # is there a temp file for this lemma/pos combination?
+    unless @lemma2temp[lemmapos]
+      @lemma2temp[lemmapos] = Tempfile.new("fred_features")
+      @lemma2temp[lemmapos].close()
+    end
+
+    # is there an open temp file for this lemma/pos combination?
+    pair = @writers.assoc(lemmapos)
+    if pair
+      return pair.last()
+    end
+
+    # no: open the temp file, kick some other temp file out of the
+    # @writers list
+    writer = @lemma2temp[lemmapos]
+    writer.open()
+    
+
+    # writer: open for appending
+    writer.seek(0, IO::SEEK_END)
+   
+
+    @writers << [lemmapos, writer]
+    if @writers.length() > @size
+      # close file associated with first writer
+      @writers.first.last.close()
+      @writers.shift()
+    end
+    return writer
+  end
+
+  ###
+  def remove_files()
+    @lemma2temp.each_value { |x|
+      x.close(true)
+    }
+  end
+end
+
+##############
+# write features,
+# either lemma-wise
+# or lemma+sense-wise
+# if lemma+sense-wise, write as binary classifier,
+# i.e. map the target senses 
+#
+# Use Delegator.
+
+###
+# Features for N-ary classifiers
+class WriteFeaturesNary
+  def initialize(lemma,
+		 exp,
+		 dataset,
+		 feature_dir)
+
+    @filename = feature_dir + fred_feature_filename(lemma)
+    @f = File.new(@filename, "w")
+    @handle_multilabel = exp.get("handle_multilabel")
+  end
+
+  def write_instance(features, senses)
+
+    @f.print features.map { |x|
+      x.to_s().gsub(/,/, "COMMA").gsub(/;/, "SEMICOLON")
+    }.join(",")
+
+    # possibly more than one sense? then use semicolon to separate
+    if @handle_multilabel == "keep"
+      # possibly more than one sense: 
+      # separate by semicolon,
+      # and hope that the classifier knows this
+      @f.print ";"
+      @f.puts senses.map {|x|
+        x.to_s().gsub(/,/, "COMMA").gsub(/;/, "SEMICOLON")
+      }.join(",")
+    else
+      # one sense: just separate by comma
+      @f.print ","
+      @f.puts senses.first().to_s().gsub(/,/, "COMMA").gsub(/;/, "SEMICOLON")
+    end
+  end
+
+  def close()
+    @f.close()
+  end
+end
+
+###
+# Features for binary classifiers
+class WriteFeaturesBinary
+  def initialize(lemma,
+		 exp,
+		 dataset,
+		 feature_dir)
+    @dir = feature_dir
+    @lemma = lemma
+    @feature_dir = feature_dir
+
+    @negsense = exp.get("negsense")
+    unless @negsense
+      @negsense = "NONE"
+    end
+
+    # files: sense-> filename
+    @files = Hash.new()
+
+    # keep all instances such that, when a new sense comes around,
+    # we can write them for that sense
+    @instances = Array.new()
+  end
+
+
+  def write_instance(features, senses)
+    # sense we haven't seen before? Then we need to
+    # write the whole featurization file for that new sense
+    check_for_presence_of_senses(senses)
+
+    # write this new instance for all senses
+    @files.each_key { |sense_of_file|
+      write_to_sensefile(features, senses, sense_of_file)
+    }
+
+    # store instance in case another sense crops up later
+    @instances << [features, senses]
+  end
+
+
+  ###
+  def close()
+    @files.each_value { |f| f.close() }
+  end
+
+  ######
+  private
+
+  def check_for_presence_of_senses(senses)
+    senses.each { |sense|
+      # do we have a sense file for this sense?
+      unless @files[sense]
+        # open new file for this sense
+        @files[sense] = File.new(@feature_dir + fred_feature_filename(@lemma, sense, true), "w")
+        # filename = @feature_dir + fred_feature_filename(@lemma, sense, true)
+        # $stderr.puts "Starting new feature file #{filename}"
+
+        # and re-write all previous instances for it
+        @instances.each { |prev_features, prev_senses|
+          write_to_sensefile(prev_features, prev_senses, 
+                             sense)
+        }
+      end
+    }
+  end
+
+  ###
+  def write_to_sensefile(features, senses,
+			 sense_of_file)
+    # file to write to
+    f = @files[sense_of_file]
+
+    # print features
+    f.print features.map { |x|
+      x.to_s().gsub(/,/, "COMMA")
+    }.join(",")
+
+    f.print ","
+
+    # binarize target class
+    if senses.include? sense_of_file
+      # $stderr.puts "writing POS #{sense_of_file}"
+      f.puts sense_of_file.to_s()
+    else
+      # $stderr.puts "writing NEG #{negsense}"
+      f.puts @negsense
+    end
+
+  end
+end
+
+########
+# class writing features:
+# delegating to either a binary or an n-ary writer
+class WriteFeaturesNaryOrBinary < SimpleDelegator
+  ###
+  def initialize(lemma, 
+		 exp,
+		 dataset)
+    feature_dir = WriteFeaturesNaryOrBinary.feature_dir(exp, dataset, "new")
+    if exp.get("binary_classifiers")
+      # binary classifiers
+      # $stderr.puts "Writing binary feature data."
+
+      # delegate writing to the binary feature writer
+      @writer = WriteFeaturesBinary.new(lemma, exp, dataset, feature_dir)
+      super(@writer)
+
+    else
+      # n-ary classifiers
+      # $stderr.puts "Writing n-ary feature data."
+
+      # delegate writing to the n-ary feature writer
+      @writer = WriteFeaturesNary.new(lemma, exp, dataset, feature_dir)
+      super(@writer)
+    end
+  end
+
+  def WriteFeaturesNaryOrBinary.feature_dir(exp, dataset,
+                                            mode = "existing")
+    return fred_dirname(exp, dataset, "features", mode)
+  end
+
+  ###
+  def WriteFeaturesNaryOrBinary.remove_files(exp, dataset)
+    feature_dir = WriteFeaturesNaryOrBinary.feature_dir(exp, dataset, "new")
+
+    Dir[feature_dir + fred_feature_filename("*")].each { |filename|
+      if File.exists? filename
+	File.delete(filename)
+      end
+    }
+  end
+end