shalmaneser-fred 1.2.0.rc4 → 1.2.rc5

data/lib/fred/FredDetermineTargets.rb

@@ -1,319 +0,0 @@
-require "fred/FileZipped"
-
-require "fred/fred_config_data"
-require "common/SynInterfaces"
-require "fred/FredConventions"
-
-
-########################################
-# target determination classes:
-# either determine targets from existing annotation
-# with frames,
-# or use all known targets.
-class Targets
-  attr_reader :targets_okay
-
-  ###
-  def initialize(exp,                 # experiment file object
-                 interpreter_class,   # SynInterpreter class, or nil
-                 mode)                # string: "r", "w", "a", as in files
-    @exp = exp
-    @interpreter_class = interpreter_class
-
-    # keep recorded targets here.
-    # try to read old list now.
-    @targets = Hash.new()
-
-    # write target info in the classifier directory.
-    # This is _not_ dependent on a potential split ID
-    @dir = File.new_dir(fred_classifier_directory(@exp), "targets")
-
-    @targets_okay = true
-    case mode
-    when "w"
-      # start from scratch, no list of targets
-    when "a", "r"
-      # read existing file containing targets
-      begin
-        file = FileZipped.new(@dir + "targets.txt.gz")
-      rescue
-        # no pickle present: signal this
-        @targets_okay = false
-        return
-      end
-      file.each { |line|
-        line.chomp!
-        if line =~ /^LEMMA (.+) SENSES (.+)$/
-          lemmapos = $1
-          senses = $2.split()
-          lemmapos.gsub!(/ /, '_')
-          #lemmapos.gsub!(/\.[A-Z]\./, '.')
-         @targets[lemmapos] = senses
-        end
-      }
-
-    else
-      $stderr.puts "Error: shouldn't be here."
-      exit 1
-    end
-
-    if ["w", "a"].include? mode
-      @record_targets = true
-    else
-      @record_targets = false
-    end
-  end
-
-  ###
-  # determine_targets:
-  # for a given SalsaTigerSentence,
-  # determine all targets,
-  # each as a _single_ main terminal node
-  #
-  # We need a single terminal node in order
-  # to compute the context window
-  #
-  # returns:
-  #  hash: target_IDs -> list of senses
-  #   where target_IDs is a pair [list of terminal IDs, main terminal ID]
-  #  
-  #  where a sense is represented as a hash:
-  #  "sense": sense, a string
-  #  "obj":   FrameNode object
-  #  "all_targets": list of node IDs, may comprise more than a single node
-  #  "lex":   lemma, or multiword expression in canonical form
-  #  "sid": sentence ID
-  def determine_targets(sent)
-    raise "overwrite me"
-  end
-
-  ##
-  # returns a list of lemma-pos combined strings
-  def get_lemmas()
-    return @targets.keys()
-  end
-
-  ##
-  # access to lemmas and POS, returns a list of pairs [lemma, pos] (string*string)
-  def get_lemma_pos()
-
-    return @targets.keys().map { |lemmapos| fred_lemmapos_separate(lemmapos) }
-  end
-
-  ##
-  # access to senses
-  def get_senses(lemmapos) # string, result of fred_lemmapos_combine 
-    @targets[lemmapos] ? @targets[lemmapos] : []
-  end
-
-  ##
-  # write file
-  def done_reading_targets()
-    begin
-      file = FileZipped.new(@dir + "targets.txt.gz", "w")
-    rescue
-      $stderr.puts "Error: Could not write file #{@dir}targets.txt.gz"
-      exit 1
-    end
-
-    @targets.each_pair { |lemma, senses|
-      file.puts "LEMMA #{lemma} SENSES "+ senses.join(" ")
-    }
-
-    file.close
-  end
-
-  ###############################
-  protected
-  
-  ##
-  # record: record occurrence of a lemma/sense pair
-  # <@targets> data structure
-  def record(target_info) 
-    lemmapos = fred_lemmapos_combine(target_info["lex"], target_info["pos"])
-    unless @targets[lemmapos]
-      @targets[lemmapos] = []
-    end
- 
-    unless @targets[lemmapos].include? target_info["sense"]
-      @targets[lemmapos] << target_info["sense"]
-    end
-  end
-end
-
-########################################
-class FindTargetsFromFrames < Targets
-  ###
-  # determine_targets:
-  # use existing frames to find targets
-  #
-  # returns:
-  #  hash: target_IDs -> list of senses
-  #   where target_IDs is a pair [list of terminal IDs, main terminal ID]
-  #  
-  #  where a sense is represented as a hash:
-  #  "sense": sense, a string
-  #  "obj":   FrameNode object
-  #  "all_targets": list of node IDs, may comprise more than a single node
-  #  "lex":   lemma, or multiword expression in canonical form
-  #  "sid": sentence ID
-  def determine_targets(st_sent) #SalsaTigerSentence object
-   retv = Hash.new()
-    st_sent.each_frame { |frame_obj|
-      # instance-specific computation:
-      # target and target positions
-      # WARNING: at this moment, we are 
-      # not considering true multiword targets for German.
-      # Remove the "no_mwe" parameter in main_node_of_expr
-      # to change this
-      term = nil
-      all_targets = nil
-      if frame_obj.target.nil? or frame_obj.target.children.empty?
-        # no target, nothing to record
-
-      elsif @exp.get("language") == "de"
-        # don't consider true multiword targets for German
-        all_targets = frame_obj.target.children()
-        term = @interpreter_class.main_node_of_expr(all_targets, "no_mwe")
-
-      else
-        # for all other languages: try to figure out the head target word
-        # anyway
-        all_targets = frame_obj.target.children()
-        term = @interpreter_class.main_node_of_expr(all_targets)
-      end
-
-      if term and term.is_splitword?
-        # don't use parts of a word as main node
-          term = term.parent()
-      end
-      if term and term.is_terminal?
-        key = [all_targets.map { |t| t.id() }, term.id()]
-
-        unless retv[key]
-          retv[key] = Array.new()
-        end
-
-        pos = frame_obj.target().get_attribute("pos")
-        # gold POS available, may be in wrong form,
-        # i.e. not the same strings that @interpreter_class.category()
-        # would return
-        case pos
-        when /^[Vv]$/
-          pos = "verb"
-        when /^[Nn]$/
-          pos = "noun"
-        when /^[Aa]$/
-          pos = "adj"
-        when nil
-          pos = @interpreter_class.category(term)
-        end
-
-        target_info = {
-          "sense" => frame_obj.name(),
-          "obj" => frame_obj,
-          "all_targets" => frame_obj.target.children().map { |ch| ch.id() },
-          "lex" => frame_obj.target().get_attribute("lemma"),
-          "pos" => pos,
-          "sid" => st_sent.id()
-        }
-      #print "lex ", frame_obj.target(), " und ",frame_obj.target().get_attribute("lemma"), "\n"
-        retv[key] << target_info
-        if @record_targets
-          record(target_info)
-        end
-      end
-    }
-    return retv
-  end
-end
-
-########################################
-class FindAllTargets < Targets
-  ###
-  # determine_targets:
-  # use all known lemmas, minus stopwords
-  def initialize(exp,
-                 interpreter_class)
-    # read target info from file
-    super(exp, interpreter_class, "r")
-    @training_lemmapos_pairs = get_lemma_pos()
-
-    get_senses(@training_lemmapos_pairs)
-    # list of words to exclude from assignment, for now
-    @stoplemmas = [
-                   "have", 
-                   "do", 
-                   "be"
-                   #      "make"
-                  ]
-  
-  end
-
-  ####
-  #
-  # returns:
-  #  hash: target_IDs -> list of senses
-  #   where target_IDs is a pair [list of terminal IDs, main terminal ID]
-  #  
-  #  where a sense is represented as a hash:
-  #  "sense": sense, a string
-  #  "obj":   FrameNode object
-  #  "all_targets": list of node IDs, may comprise more than a single node
-  #  "lex":   lemma, or multiword expression in canonical form
-  #  "sid": sentence ID
-  def determine_targets(sent) #SalsaTigerSentence object
-    # map target IDs to list of senses, in our case always [ nil ]
-    # because we assume that the senses of the targets we point out
-    # are unknown
-    retv = Hash.new()
-    # iterate through terminals of the sentence, check for inclusion
-    # of their lemma in @training_lemmas
-    sent.each_terminal { |node|
-      # we know this lemma from the training data,
-      # and it is not an auxiliary,
-      # and it is not in the stopword list
-      # and the node does not represent a preposition
-
-      ### modified by ines, 17.10.2008
-      lemma = @interpreter_class.lemma_backoff(node)
-      pos = @interpreter_class.category(node)
-
-#	print "lemma ", lemma, " pos ", pos, "\n"
-#      reg = /\.[ANV]/
-#      if !reg.match(lemma) 
-#        if /verb/.match(pos) 
-#          lemma = lemma + ".V"
-#        elsif /noun/.match(pos) 
-#          lemma = lemma + ".N"
-#        elsif /adj/.match(pos) 
-#          lemma = lemma + ".A"
-#        end
-#        print "LEMMA ", lemma, " POS ", pos, "\n" 
-#      end
-
-      if (@training_lemmapos_pairs.include? [lemma, pos] and
-          not(@interpreter_class.auxiliary?(node)) and
-          not(@stoplemmas.include? lemma) and
-          not(pos == "prep"))
-          key = [ [ node.id() ], node.id() ]
-
-          # take this as a target.
-          retv[ key ] = [ 
-                         {
-                           "sense" => nil,
-                           "obj" => nil,
-                           "all_targets" => [ node.id() ],
-                           "lex" => lemma,
-                           "pos" => pos,
-                           "sid" => sent.id()
-                         } ]
-          # no recording of target info,
-          # since we haven't determined anything new
-        end
-    }
-
-    return retv
-  end
-end
-

data/lib/fred/FredEval.rb

@@ -1,312 +0,0 @@
-# FredEval
-# Katrin Erk April 05
-#
-# Frame disambiguation system: evaluate classification results
-#
-# While the other main classes of Fred just provide a new() method
-# and a compute() method,
-# the FredEval class also provides access methods to all the
-# individual evaluation results and allows for a flag that
-# suppresses evaluation output to a file --
-# such that this package can also be used by external systems that
-# wish to evaluate Fred.
-#
-# Inherits from the Eval class that is not Fred-specific
-
-# Salsa packages
-require "common/Eval"
-require "common/ruby_class_extensions"
-
-# Fred packages
-require "fred/fred_config_data"
-require "fred/FredConventions"
-require "fred/FredFeatures"
-require "fred/FredDetermineTargets"
-
-class FredEval < Eval
-
-  ###
-  # new
-  #
-  # evaluate runtime options and announce the task
-  def initialize(exp_obj, # FredConfigData object
-		 options) # hash: runtime option name (string) => value(string)
-
-    in_enduser_mode_unavailable()
-
-    @exp = exp_obj
-
-    ###
-    # evaluate runtime options
-    @split_id = nil
-    logfilename = nil
-
-    options.each_pair { |opt, arg|
-      case opt
-      when "--logID"
-	
-	@split_id = arg
-      when "--printLog"
-	logfilename = fred_dirname(@exp, "eval", "log", "new") +
-                      "eval_logfile.txt"
-	
-      else
-	# case of unknown arguments has been dealt with by fred.rb
-      end
-    }
-
-    ###
-    # make outfile name
-    outfilename =  fred_dirname(@exp, "eval", "eval", "new") + 
-                   "eval.txt"
-
-    ###
-    # do we regard all senses as assigned,
-    # as long as they surpass some threshold?
-    # if we are doing multilabel evaluation, we need the full list of senses
-    @threshold = @exp.get("assignment_confidence_threshold")
-    @target_obj = Targets.new(@exp, nil, "r")
-    unless @target_obj.targets_okay
-      # error during initialization
-      $stderr.puts "Error: Could not read list of known targets, bailing out."
-      exit 1
-    end
-
-    if @threshold or @exp.get("handle_multilabel") == "keep"
-      @multiple_senses_assigned = true
-    else
-      @multiple_senses_assigned = false
-    end
-
-
-    ###
-    # initialize abstract class behind me
-    if @multiple_senses_assigned
-      # we are possibly assigning more than one sense: do precision/recall
-      # instead of accuracy:
-      # "true" is what "this sense has been assigned" is mapped to below.
-      super(outfilename, logfilename, "true")
-    else
-      super(outfilename, logfilename)
-    end
-
-    # what is being done with instances with multiple sense labels?
-    @handle_multilabel = @exp.get("handle_multilabel")
-
-    ###
-    # announce the task
-    $stderr.puts "---------"
-    $stderr.print "Fred  experiment #{@exp.get("experiment_ID")}: Evaluating classifiers"
-    if @split_dir
-      $stderr.puts " using split with ID #{@split_id}"
-    else
-      $stderr.puts
-    end
-    if @multiple_senses_assigned
-      $stderr.puts "Allowing for the assignment of multiple senses,"
-      $stderr.puts "computing precision and recall against the full sense list of a lemma."
-    end
-    $stderr.puts "Writing result to #{fred_dirname(@exp, "eval", "eval")}"
-    $stderr.puts "---------"
-  end
-
-  #####
-  protected
-
-  ###
-  # each_group
-  #
-  # yield each group name in turn
-  # in our case, group names are lemmas
-  #
-  # also, set object-global variables in such a way
-  # that the elements of this group can be read
-  def each_group()
-
-    # access to classifier output files
-    output_dir = fred_dirname(@exp, "output", "tab")
-    # access to answer key files
-
-    
-    if @split_id
-      # make split object and parameter hash to pass to it
-      dataset = "train"
-    else
-      dataset = "test"
-    end
-
-    # iterate through instance files
-    @target_obj.get_lemmas().sort().each { |lemma|
-      # progress report
-      if @exp.get("verbose")
-        $stderr.puts "Evaluating " + lemma
-      end
-
-      # file with classification results
-      begin
-        @classfile = File.new(output_dir + fred_result_filename(lemma))
-      rescue
-        # no classification results
-        @classfile = nil
-      end
-
-      # file with answers:
-      # maybe we need to apply a split first
-      if @split_id
-        @goldreader = AnswerKeyAccess.new(@exp, "train", lemma, "r", @split_id, "test")
-      else
-        @goldreader = AnswerKeyAccess.new(@exp, "test", lemma, "r")
-      end      
-
-      # doing multilabel evaluation?
-      # then we need a list of all senses
-      if @multiple_senses_assigned
-        @all_senses = @target_obj.get_senses(lemma)
-      else
-        @all_senses = nil
-      end
-
-      yield lemma
-    }
-  end
-
-  ###
-  # each_instance
-  #
-  # given a lemma name, yield each instance of this lemma in turn,
-  # or rather: yield pairs [gold_class(string), assigned_class(string)]
-  #
-  # relies on each_group() having set the appropriate readers
-  # <@goldreader> and <@classfile>
-  def each_instance(lemma) # string: lemma name
-
-    # watch out for repeated instances
-    # which may occur if handle_multilabel = repeat.
-    # Only yield them once to avoid re-evaluating multi-label instances
-    #
-    # instance_ids_seen: hash target_ids -> true/nil
-    instance_ids_seen = Hash.new()
-
-    # read gold file and classifier output file in parallel
-    @goldreader.each { |lemma, pos, target_ids, sid, senses_gold, transformed_gold_senses|
-
-      # classline: format
-      # (label confidence)*
-      # such that the label with the highest confidence is first
-      classline = nil
-      if @classfile
-        classline = @classfile.gets()
-      end
-      if classline.nil?
-	classline = ""
-      end
-
-      # $stderr.puts "HIER0 #{classline} #{@classfile.nil?}"
-
-      # have we done this same instance previously?
-      if instance_ids_seen[target_ids]
-        next
-      end
-      # instance not seen previously, but mark as seen now.
-      instance_ids_seen[target_ids] = true
-
-      # determine all assigned senses and their confidence levels
-      # determine all sense/confidence pairs
-      # senses assigned: list of pairs [senselist, confidence]
-      # where senselist is an array of sense strings
-      senses_assigned = Array.new()
-      current_sense = nil
-
-      classline.split().each_with_index { |entry, index|
-        if index % 2 == 0
-          # we have a sense label
-          if @handle_multilabel == "join"
-            # split up joined senses
-            current_sense = fred_split_sense(entry)
-          else
-            current_sense = [entry]
-          end
-
-        else
-          # we have a confidence level
-          senses_assigned << [current_sense, entry.to_f()]
-        end
-      }
-
-
-      if @threshold
-        # multiple senses assigned, and
-        # regard as assigned everything above a given threshold
-
-        # transform senses_assigned: 
-        # in the case of "join", one sense may have several confidence levels,
-        # one on its own and one in a joined sense
-        senses_assigned_hash = Hash.new()
-        senses_assigned.each { |senses, confidence|
-          senses.each { |s|
-            # assign to each sense the maximum of its previous confidence
-            # and this one.
-            # watch out: confidence may be smaller than zero
-            if senses_assigned_hash[s]
-              senses_assigned_hash[s] = [senses_assigned_hash[s], confidence].max()
-            else
-              senses_assigned_hash[s] = confidence
-            end
-          }
-        }
-
-        # select all sense/confidence pairs where confidence is above threshold
-        senses_assigned = senses_assigned_hash.to_a().select { |sense, confidence|
-          confidence >= @threshold
-        }.map { |sense, confidence| 
-          # then retain only the sense, not the confidence
-          sense 
-        }
-
-
-        unless @all_senses
-          raise "Shouldn't be here"
-        end
-
-        # for each sense out of the list of all senses:
-        # yield a pair of [applies, has been assigned]
-        # both 'applies' and 'has been assigned' will be
-        # a string of either 'true' or 'false'
-        # assignment is accurate if both are the same
-        @all_senses.each { |sense_of_lemma|
-          gold_class = (senses_gold.include? sense_of_lemma).to_s()
-          assigned_class = (senses_assigned.include? sense_of_lemma).to_s()
-          yield [gold_class, assigned_class]
-        }
-
-
-      else
-        # regard only one sense as assigned at a time
-        # count as correct if the list of gold classes
-        # contains the main assigned class
-        # (relatively lenient evaluation)
-
-        # actually assigned class: only the one with the 
-        # maximum confidence
-        # $stderr.puts "HIER5 #{senses_assigned.length()}"
-
-        if senses_assigned.empty?
-          # nothing to yield
-        else
-
-          max_senselist = senses_assigned.max { |a, b|
-            a.last() <=> b.last()
-          }.first()
-          
-
-          max_senselist.each { |single_sense|
-            gold_class = (senses_gold.include? single_sense).to_s()
-            yield [gold_class, "true"]
-          }
-        end
-
-      end
-    }
-  end
-
-end