shalmaneser 0.0.1.alpha

data/lib/rosy/GfInduce.rb

@@ -0,0 +1,1115 @@
+# GfInduce
+# Katrin Erk Jan 2006
+#
+# Given parse trees with FrameNet frames assigned on top of the syntactic analysis,
+# and given that the Frame Elements also contain information on grammatical function
+# and phrase type (as e.g. in the FrameNet annotation),
+# induce a mapping from parse tree paths to grammatical functions from this information
+# and apply it to new sentences
+
+require "common/AbstractSynInterface"
+require "common/ruby_class_extensions"
+
+#####################################################################
+# Management of mapping from GFs to paths
+#####################################################################
+
+class GfiGfPathMapping
+
+  #########################################
+  # Initialization
+  #########################################
+
+  ###
+  def initialize(interpreter_class)
+
+    @interpreter = interpreter_class
+
+    # hash: POS(string) -> hash gf(string) -> hash: path_string -> frequency(int)
+    @gf_to_paths = Hash.new
+
+    # hash: POS(string)-> hash: gf(string) -> hash: one edge of a path -> 
+    #  frequency(int) | hash: one edge of a path -> ...
+    @gf_to_edgelabel = Hash.new
+
+    # hash: word(string) -> array: [gf, prep, head_category]
+    @word_to_gflist = Hash.new
+
+    # hash: path as string(string) -> array of steps
+    # where a step is a tuple of stringss [{U, D}, edgelabel, nodelabel}
+    @pathstring_to_path = Hash.new
+  end
+
+  #########################################
+  # Storing induced mappings
+  #########################################
+
+  ###
+  def store_mapping(gf,   # grammatical function: string
+                    path, # Path object (from AbstractSynInterface)
+                    node, # SynNode associated with GF and reached via path
+                    lemma,# lemma: string
+                    pos)  # part of speech: string
+
+    path_s = path.print(true, true, true)
+    lemmapos = string_lemmapos(lemma, pos)
+    prep = @interpreter.preposition(node)
+    if prep
+      prep.downcase!
+    end
+    h = @interpreter.head_terminal(node)
+    if h
+      headcat = @interpreter.category(h)
+    else
+      headcat = nil
+    end
+
+    # remember the path as an array of triples [direction, edgelabel, nodelabel]
+    # as hash value of the path-as-string
+    unless @pathstring_to_path[path_s]
+      @pathstring_to_path[path_s] = Array.new
+      path.each_step { |direction, edgelabel, nodelabel, node|
+        @pathstring_to_path[path_s] << [direction, edgelabel, nodelabel]
+      }
+    end
+
+    # store the mapping in the
+    # gf -> path hash
+    unless @gf_to_paths[pos]
+      @gf_to_paths[pos] = Hash.new
+    end
+    unless @gf_to_paths[pos][gf]
+      @gf_to_paths[pos][gf] = Hash.new(0)
+    end
+    @gf_to_paths[pos][gf][path_s] = @gf_to_paths[pos][gf][path_s] + 1
+
+
+    # remember this gf/pt tuple as possible GF of the current lemma
+    unless @word_to_gflist[lemmapos]
+      @word_to_gflist[lemmapos] = Array.new
+    end
+    unless @word_to_gflist[lemmapos].include? [gf, prep, headcat]
+      @word_to_gflist[lemmapos] << [gf, prep, headcat]
+    end
+  end
+
+  ###
+  # finish up inducing:
+  #  reencode information in a fashion
+  #  that makes apply() faster
+  def finish_inducing()
+    # make sure gf_to_edgelabel is empty at the start
+    @gf_to_edgelabel.clear()
+    
+    @gf_to_paths.each_pair { |pos, gf_to_paths_to_freq|
+      unless @gf_to_edgelabel[pos]
+        @gf_to_edgelabel[pos] = Hash.new()
+      end
+
+      gf_to_paths_to_freq.each_pair { |gf, paths_to_freq|
+        paths_to_freq.each_pair { |pathstring, freq|
+
+          steps = @pathstring_to_path[pathstring]
+          if steps.nil? or steps.empty?
+            # do not list empty paths
+            $stderr.puts "found empty path for #{gf}, frequency #{freq}. Skipping."
+            next
+          end
+        
+          if freq >= 5 or
+              gf =~ /Head|Appositive|Quant|Protagonist/
+            # path frequent enough: list it
+
+            unless @gf_to_edgelabel[pos][gf]
+              @gf_to_edgelabel[pos][gf] = Hash.new()
+            end
+            
+            enter_path(@gf_to_edgelabel[pos][gf], steps.clone(), freq)
+          end
+        }
+      }
+    }
+  end
+
+  #########################################
+  # Test output
+  #########################################
+
+  ###
+  # test output
+  def test_output()
+    # gf_to_paths:
+    # sum frequencies, compare frequency against average path length
+    puts "============================="
+    puts "GF_TO_PATHS"
+    puts "============================="
+#     @gf_to_paths.each_key { |pos|
+#       @gf_to_paths[pos].each_key { |gf|
+#         puts "================"
+#         puts "POS #{pos} GF #{gf}:"
+#         @gf_to_paths[pos][gf].each_pair { |path_s, freq|
+#           puts "#{path_s} freq:#{freq} len:#{@pathstring_to_path[path_s].length()}"
+#         }
+#       }
+#     }  
+    @gf_to_paths.each_key { |pos|
+      @gf_to_paths[pos].each_key { |gf|
+        puts "================"
+        puts "POS #{pos} GF #{gf}:"
+  
+        @gf_to_paths[pos][gf].values.uniq.sort { |a, b| b <=> a}.each { |frequency|
+          sum = 0
+          count = 0
+          @gf_to_paths[pos][gf].each_pair { |path_s, otherfreq|
+            if otherfreq == frequency
+              count += 1
+              sum += @pathstring_to_path[path_s].length()
+            end
+          }
+          avg_pathlen = sum.to_f / count.to_f
+          
+          puts "     Frequency #{frequency}: #{count} path(s)"
+          puts "                  #{avg_pathlen} avg. path len"
+        }
+        puts
+      }
+    }
+    puts
+    puts "WORD_TO_GFLIST"
+    puts "============================="
+    @word_to_gflist.each_pair { |word, gflist|
+      print word, " ", gflist.map { |gf, prep, hc| "GF:[#{gf}] PREP:#{prep} HEADCAT:#{hc}" }.join(", "), "\n"
+    }
+    puts
+    puts "============================="
+    puts "GF TO EDGELABEL"
+    puts "============================="
+    @gf_to_edgelabel.each_key { |pos|
+      @gf_to_edgelabel[pos].each_pair { |gf, entries|
+        puts "POS #{pos} GF #{gf}"
+        print_entries(entries, 2)
+      }
+    }
+  end
+
+  #########################################
+  # Restricting induced mappings 
+  # to achieve better mappings
+  #########################################
+
+  ####
+  # restrict gf_to_edgelabel hashes:
+  # exclude all paths that include an Up edge
+  #
+  # changes @gf_to_edgelabel, not reversible
+  def restrict_to_downpaths()
+    @gf_to_edgelabel.each_value { |pos_specific|
+      pos_specific.each_value { |hash_or_val|
+        restrict_pathhash_to_downpaths(hash_or_val)
+      }
+    }
+  end
+
+  ####
+  # restrict gf_to_edgelabel hashes:
+  # only keep paths up to length n
+  #
+  # changes @gf_to_edgelabel, not reversible
+  def restrict_pathlen(n)  # integer: maximum length to keep
+    @gf_to_edgelabel.each_value { |pos_specific|
+      pos_specific.each_value { |hash_or_val|
+        restrict_pathhash_len(hash_or_val, n)
+      }
+    }
+  end
+
+  ####
+  # restrict gf_to_edgelabel hashes:
+  # remove GFs that are often incorrect
+  def remove_gfs(gf_list)
+    gf_list.each { |gf|
+      # test output
+      @gf_to_edgelabel.each_value { |pos_specific|
+        if pos_specific[gf]
+#          puts "Remove GFs: removing #{gf}"
+        end
+        pos_specific.delete(gf)
+      }
+    }
+  end
+
+  #########################################
+  # Using stored data
+  #########################################
+
+
+  ###
+  # given a SynNode,
+  # return all its potential GFs
+  # by comparing paths in the parse tree
+  # against the GF/path mappings stored in @gf_to_edgelabel
+  #
+  # returns: 
+  # hash: SynNode -> tuple [GF(string), preposition(string), frequency(integer)]
+  def potential_gfs_of_node(start_node,  # SynNode
+                            lemma,       # string: lemma for start_node
+                            pos)
+
+
+    # determine possible GFs of a SynNode:
+    # 
+    # hash: SynNode(some node in this sentence) -> list of tuples [gf label, prep, headcat, hash of steps]
+    # initialize with just the entry for the start node
+    potential_gfs = Hash.new
+    potential_gfs[start_node] = potential_gfs_of_lemma(lemma, pos)
+#     $stderr.puts "HIER #{lemma} " + potential_gfs_of_lemma(lemma, pos).map { |gf, prep, hc, hash|
+#       "#{gf}:#{prep}:#{hc} "
+#     }.join(" ")
+
+    # agenda: list of SynNode objects
+    #  that have been considered as potential GFs in the previous step
+    #  next: consider their surrounding nodes
+    #
+    # so, we never assign a GF to the start node
+    agenda = [start_node]
+    # been_there: list of SynNode objects
+    #  that have been considered already and needn't be visited again
+    been_there = Hash.new
+    been_there[start_node] = true
+
+    # hash: SynNode -> tuple [GF(string), preposition(string), frequency(integer)]
+    #      node identified for this sentence for GF,
+    #      frequency: frequency with which the path from verb to GF has
+    #                 been seen in the FN data (such that we can keep
+    #                 the best path and discard others)
+    node_to_label_and_freq = Hash.new()
+
+    while not(agenda.empty?)
+      prev_node = agenda.shift()
+
+      unless potential_gfs[prev_node]
+        # no further GFs to be reached from prev_node:
+        # shouldn't be here, but never mind, just ignore
+        next
+      end
+      
+      # surrounding_nodes returns a list of pairs [SynNode, Path object]
+      @interpreter.surrounding_nodes(prev_node, true).each { |node, path|
+        myprep = @interpreter.preposition(node)
+        if myprep
+          myprep.downcase!
+        end
+        h = @interpreter.head_terminal(node)
+        if h
+          my_headcat = @interpreter.category(h)
+        else
+          my_headcat = nil
+        end
+
+        if been_there[node]
+          next
+        end
+
+        been_there[node] = true
+
+        unless potential_gfs[node]
+          potential_gfs[node] = Array.new
+        end
+      
+        path.each_step() { |step|
+          # each edge from prev_node to node:
+          # see whether we can walk this edge to reach some of the GFs
+          # still to be reached
+
+          step_s = string_step(step)
+
+          potential_gfs[prev_node].each { |gf, prep, headcat, hash|
+
+            if hash[step_s]
+              # yes, there is still a possibility of reaching gf 
+              # from our current node
+
+              if hash[step_s].kind_of? Integer
+                # actually, we have reached gf,
+                # and hash[last_edge] is the frequency with which
+                # this path has led to this GF in the FN data
+
+                freq = hash[step_s]
+                
+                # check whether node has the right preposition
+                # and the right head category
+                if myprep != prep or
+                    my_headcat != headcat
+                  # we were supposed to find a preposition
+                  # but didn't , or didn't find the right one;
+                  # or we got the wrong head category
+                  # discard current entry
+
+                elsif not(node_to_label_and_freq[node]) or
+                    node_to_label_and_freq[node].last < freq
+                  # this node has not been assigned any GF before,
+                  # or the old frequency was lower than the current one:
+                  # keep the new entry
+                  node_to_label_and_freq[node] = [gf, prep, freq]
+                  
+                else
+                  # this node has been assigned a GF before, and the 
+                  # other frequency was higher:
+                  # discard the current entry
+                end
+                
+              else
+                # we have not yet reached gf, but we still might
+                # at the next node we meet from here
+                potential_gfs[node] << [gf, prep, headcat, hash[step_s]]
+              end
+            end
+          } # each gf/hash pair for prev_node
+        } # each edge leading from prev_node to node
+
+        # further explore the parse from this node?
+        # only if there are still GFs to be reached from here
+        unless potential_gfs[node].empty?
+          unless agenda.include? node
+            agenda << node
+          end
+        end
+      } # each surrounding node of prev_node
+    end # while agenda nonempty
+
+    return node_to_label_and_freq
+  end
+
+
+
+  ####################################
+  ####################################
+  private
+
+  #########################################
+  # Strings for hashing
+  #########################################
+
+  def string_lemmapos(lemma, pos)
+    return lemma.to_s + "!" + pos.to_s
+  end
+
+  ###
+  # make key for gf_to_edgelabel hash
+  #
+  # step: array of things, the first 3 being strings
+  #      direction, edgelabel, nodelabel
+  #
+  # returns: string, the key
+  def string_step(step) 
+    direction = step[0]
+    edgelabel = step[1]
+    nodelabel = step[2]
+
+    return "#{direction} #{edgelabel} #{nodelabel}"
+  end
+
+  #########################################
+  # Storing induced mappings
+  #########################################
+
+  ####
+  # build up linked hashes that map
+  # paths to frequencies
+  def enter_path(hash,       # partial result of enter_path
+                 chainlinks, # array: string*string*string
+                 frequency)  # integer: frequency of this mapping
+    # take off first chain link
+    key = string_step(chainlinks.shift())
+    
+    if chainlinks.empty?
+      # that was the last link, actually
+      hash[key] = frequency
+    else
+      # more links available
+      unless hash[key]
+        hash[key] = Hash.new()
+      end
+      
+      if hash[key].kind_of? Integer
+        # there is a shorter path for the same GF, 
+        # ending at the point where we are now.
+        # which frequency is higher?
+        if frequency > hash[key]
+          hash[key] = Hash.new()
+        else
+          return
+        end
+      end
+      
+      enter_path(hash[key], chainlinks, frequency)
+    end
+  end
+
+
+  #########################################
+  # Test output
+  #########################################
+
+  ###
+  # test output:
+  # print results of enter_path
+  def print_entries(hash, num_spaces)
+    hash.each_pair { |first_link, rest|
+      print " "*num_spaces, first_link
+      
+      if rest.kind_of? Integer
+        puts "  #{rest}"
+      else
+        puts
+        print_entries(rest, num_spaces + 2)
+      end
+    }
+  end
+
+  #########################################
+  # Restricting induced mappings 
+  # to achieve better mappings
+  #########################################
+
+  ###
+  # recursive function:
+  # if the argument is a hash,
+  # kill all entries whose keys describe an Up step in the path,
+  # go into recursion for remaining entries
+  def restrict_pathhash_to_downpaths(hash_or_val) # path hash or integer freq
+    if hash_or_val.kind_of? Integer
+      return
+    end
+
+    # remove up edges
+    hash_or_val.delete_if { |key, val|
+      # test output
+#      if key =~ /^U/
+#        puts "Deleting up path"
+#      end
+      key =~ /^U/
+    }
+
+    hash_or_val.each_value { |next_hash|
+      restrict_pathhash_to_downpaths(next_hash)
+    }
+  end
+
+  ###
+  # recursive function:
+  # if the argument is a hash and 
+  # the remaining path length is 0, kill all entries
+  # else go into recursion for all entries with reduced path length
+  def restrict_pathhash_len(hash_or_val,  # path hash or integer freq
+			    n)            # restrict paths from what length?
+    if hash_or_val.kind_of? Integer
+      return
+    end
+
+    if n == 0
+      # test output
+ #     hash_or_val.keys.each { |k| puts "deleting because of path len: #{k}" }
+      hash_or_val.keys.each { |k| hash_or_val.delete(k) }
+    else
+      hash_or_val.each_value { |next_hash|
+	restrict_pathhash_len(next_hash, n-1)
+      }
+    end
+  end
+
+  #########################################
+  # Using stored data
+  #########################################
+
+  ###
+  # given a lemma,
+  # look in its list of all GFs that we have ever found for that lemma
+  # 
+  # returns: array of pairs [gf label, point in gf_to_edgelabel hash]
+  #   all the labels of GFs of this word, 
+  #   and for each GF, the matching GF-to-path hash
+  def potential_gfs_of_lemma(lemma, pos)
+
+    lemmapos = string_lemmapos(lemma, pos)
+
+    if @word_to_gflist[lemmapos]
+      return @word_to_gflist[lemmapos].map { |gf, prep, headcat|
+        [gf, prep, headcat, @gf_to_edgelabel[pos][gf]]
+      }.select { |gf, prep, headcat, hash|
+#         if hash.nil?
+#           $stderr.puts "Mapping words to GF lists: no entry for GF >>#{gf}<< for POS #{pos}"
+#         end
+        not(hash.nil?)
+      }
+    else
+      return []
+    end
+  end
+end
+
+#####################################################################
+# class managing subcat frames
+#####################################################################
+
+
+class GfiSubcatFrames
+
+  #########################################
+  # Initialization
+  #########################################
+
+  ###
+  # include_sem: include frame and FE names in
+  # subcat frame? if not, the tuple arity stays the same,
+  # but frame and FE entries will be nil
+  def initialize(include_sem) # boolean
+    # hash: word(string) -> array:[frame(string), subcatframe]
+    #  with subcatframe an array of tuples [gf, prep, fe, multiplicity]
+    @word_to_subcatframes = Hash.new
+
+    # hash: <subcatframe encoded as string> -> frequency
+    @subcat_to_freq = Hash.new(0)
+
+    @include_sem = include_sem
+  end
+
+  #########################################
+  # Storing induced mappings
+  #########################################
+
+  ###
+  # store a subcat frame in this object.
+  # subcat frame given as an array of tuples
+  #  [gf, prep, fe]
+  def store_subcatframe(scf,   # tuples as described above
+                        frame, # frame: string
+                        lemma, # lemma: string
+                        pos)   # part of speech: string
+
+    lemmapos = string_lemmapos(lemma, pos)
+    unless @include_sem
+      frame = nil
+    end
+    
+    unless @word_to_subcatframes[lemmapos]
+      @word_to_subcatframes[lemmapos] = Array.new
+    end
+
+    # reencode subcat frame:
+    # array of tuples [gf, prep, fe_concat, multiplicity]
+    #
+    # multiplicity is either "one" or "many", depending on
+    # the number of times the same gf/prep pair occurred.
+    # If the same gf/prep pair occurred with different FEs, they 
+    # will be concatenated into a space-separated string
+    # with a single subcat entry
+    count_gfprep = Hash.new(0)
+    gfprep_to_fe = Hash.new
+
+    scf.each { |gf, prep, fe|
+      count_gfprep[[gf, prep]] += 1
+	unless gfprep_to_fe[[gf, prep]]
+	  gfprep_to_fe[[gf, prep]] = Array.new
+	end
+	unless gfprep_to_fe[[gf, prep]].include?(fe)
+	  gfprep_to_fe[[gf, prep]] << fe
+	end
+    }
+    subcatframe = count_gfprep.to_a.map { |gfprep, count|
+      gf, prep = gfprep
+      if @include_sem
+	fe = gfprep_to_fe[[gf, prep]].join(" ")
+      else
+	fe = nil
+      end
+      if count == 1
+        [gf, prep, fe, "one"]
+      else
+        [gf, prep, fe, "many"]
+      end
+    }.sort { |a, b|
+      if a[0] != b[0]
+        # compare GF
+        a[0] <=> b[0]
+      else 
+        # compare prep
+        a[1].to_s <=> b[1].to_s
+      end
+    }
+
+    # store subcat frame
+    unless @word_to_subcatframes[lemmapos].include? [frame, subcatframe]
+      @word_to_subcatframes[lemmapos] << [frame, subcatframe]
+    end
+
+    # count subcat frame
+    @subcat_to_freq[string_subcatframe(subcatframe)] += 1
+  end
+
+  #########################################
+  # Test output
+  #########################################
+
+  ###
+  def test_output()
+    puts "WORD_TO_SUBCATFRAMES"
+    @word_to_subcatframes.each_pair { |word, frames_and_mappings|
+      puts word
+      frames_and_mappings.each { |frame, subcatframe|
+        puts "\t#{frame} "+ subcatframe.to_a.map { |gf, prep, fe, freq| "[#{gf}]:#{prep}:#{fe}:#{freq}" }.join(" ")
+	puts "\t\tfreq #{@subcat_to_freq[string_subcatframe(subcatframe)]}"
+      }
+    }
+    puts
+  end
+
+  #########################################
+  # Using stored data
+  #########################################
+
+  ###
+  def lemma_known(lemma, pos) # string*string
+    if @word_to_subcatframes[string_lemmapos(lemma, pos)]
+      return true
+    else
+      return false
+    end
+  end
+
+
+  ###
+  # given a mapping from nodes to gf/prep pairs,
+  # match them against the subcat frames known for the lemma/POS pair.
+  #
+  # node_to_gf:
+  # hash: SynNode -> tuple [GF(string), preposition(string), frequency(integer)]
+  #
+  # strict: boolean. If true, return only those subcat frames that exactly match
+  #   all GFs listed in node_to_gf. If false, also return subcat frames that 
+  #   match a subset of the GFs mentioned in node_to_gf.
+  # 
+  # returns: list of tuples [frame(string), subcat frame, frequency(integer)], 
+  # where a subcat frame is an array of tuples 
+  # [gf (string), prep(string or nil), fe(string), synnodes(array:SynNode)]
+  #    and the syn_nodes are sorted by confidence, best first
+  def match(start_node, # SynNode
+            lemma,      # string
+            pos,         # string
+            node_to_gf, # hash as described above
+            strict)     # boolean: true: strict match. false: subseteq match
+
+    unless lemma_known(lemma, pos)
+      return []
+    end
+
+#     $stderr.puts "HIER4 GFs found: " + node_to_gf.values.map { |gf, prep, freq| 
+#       "#{gf}:#{prep}" 
+#     }.join(" ")
+#     $stderr.puts "HIER5 GF possible: (#{@word_to_subcatframes[string_lemmapos(lemma, pos)].length()})"
+#     @word_to_subcatframes[string_lemmapos(lemma, pos)].each { |frame, scf|
+#       scf.each { |gf, prep, fe, mult|
+#         $stderr.print "#{gf}:#{prep} "
+#       }
+#       $stderr.puts
+#     }
+
+    # word_to_subcatframes: 
+    # hash: lemma(string) -> array:[frame(string), subcatframe]
+    #  with subcatframe: array of tuples [gf, prep, fe, multiplicity]
+    scf_list = @word_to_subcatframes[string_lemmapos(lemma, pos)].map { |frame, subcatframe|
+      [
+        frame, 
+        # returns: array of tuples [gf, prep, fe, syn_nodes]
+        match_subcat(subcatframe, node_to_gf, strict),
+        @subcat_to_freq[string_subcatframe(subcatframe)]
+      ]
+    }.select { |frame, subcatframe, frequency| not(subcatframe.nil?) }
+
+    # scf_list may contain duplicates if some GF exists both with multiplicity "many" and
+    # muiltiplicity "one", and the "many" has only been filled by one
+    #
+    # so sort by frequency, then discard duplicates using a "seen" hash
+    seen = Hash.new
+    return scf_list.sort { |a, b| b.last <=> a.last }.select { |frame, subcatframe, frequency|
+      sc_string = string_subcatframe_withnodes(subcatframe)
+      if seen[sc_string]
+        false
+      else
+        seen[sc_string] = true
+        true
+      end
+    }
+  end
+
+  ###
+  # given a subcat frame and a hash mapping each node to a gf/prep pair,
+  # check whether the node/gf mapping matches the subcat frame.
+  # Match: 
+  # * for each node/gf mapping, the GF/prep occurs in the subcat frame
+  #   (But if there are many nodes for the same GF/prep and
+  #    multiplicity is "one", nodes may be discarded.)
+  # * each entry in the subcat frame is matched by at least one node,
+  #   and multiplicity="many" entries are matched by at least two
+  #
+  # subcatframe: array of tuples [gf, prep, fe, multiplicity]
+  # node_to_gf:
+  #   hash: SynNode -> tuple [GF(string), preposition(string), frequency(integer)]
+  #
+  # returns: 
+  #  nil on mismatch.
+  #  match: copy of the subcat frame, each entry minus multiplicity but plus matching syn nodes
+  def match_subcat(subcatframe,  # array of tuples as described above
+                   node_to_gf,   # hash as described above
+                   strict)       # boolean: strict match, or subseteq match?
+
+    # each node of the node -> gf hash:
+    # check whether the GF of the node->gf mapping 
+    # occurs in the subcat frame
+    # if it does, remember it in entry_to_nodes
+    # if it does not, regard the match as failed
+    entry_to_nodes = Hash.new
+
+    node_to_gf.each_key {|node|
+      gf, prep, frequency = node_to_gf[node]
+      match_found = false
+
+      subcatframe.each { |other_gf, other_prep, fe, multiplicity|
+
+        if other_gf == gf and other_prep == prep
+          # match
+          unless entry_to_nodes[[gf, prep]]
+            entry_to_nodes[[gf, prep]] = Array.new
+          end
+          entry_to_nodes[[gf, prep]] << node
+          match_found = true
+          break
+        end
+      }
+      if strict and not(match_found)
+        # this node does not fit into this subcat frame
+        # mismatch
+        return nil
+      end
+    } # each node from node_to_gf
+
+
+    subcatframe.each { |gf, prep, fe, multiplicity|
+    
+      # opposite direction:
+      # see if all slots of the subcat frame have been matched against at least one SynNode,
+      # otherwise discard
+      unless entry_to_nodes[[gf, prep]]
+        return nil
+      end
+      
+      # only one node to be returned for this slot:
+      # use the one with the highest frequency for its gf->path mapping
+      if multiplicity == "one" and entry_to_nodes[[gf, prep]].length() > 1
+	# sort nodes by the frequency
+	# entries in node_to_gf,
+	# then keep only the <multiplicity> first ones
+	entry_to_nodes[[gf, prep]] = entry_to_nodes[[gf, prep]].sort { |node1, node2|
+	  node_to_gf[node2].last <=> node_to_gf[node1].last
+	}.slice(0, 1)
+      end
+    }
+
+    # make extended subcat frame and return it
+    return subcatframe.map { |gf, prep, fe, multiplicity|
+      # sort "many" nodes by the frequency of their gf->path mapping
+      [
+        gf, prep, fe, 
+        entry_to_nodes[[gf, prep]].sort { |node1, node2|
+          node_to_gf[node2].last <=> node_to_gf[node1].last
+        }
+      ]
+    }
+  end
+
+  ####################################
+  ####################################
+  private
+
+  #########################################
+  # Making strings for hashing
+  #########################################
+
+  ###
+  def string_lemmapos(lemma, pos)
+    return lemma.to_s + "!" + pos.to_s
+  end
+
+  ###
+  # subcatframe to string
+  #
+  # subcatframe: array of tuples [gf, prep, fe, multiplicity]
+  # sort (to make subcat frames comparable) and
+  # turn to string
+  def string_subcatframe(subcatframe)
+
+    return subcatframe.map { |gf, prep, fes, count| "#{gf} #{prep} #{count}" }.sort.join(", ")
+  end
+  
+  # subcatframe to string
+  #
+  # here: we have a list of SynNodes instead of the multiplicity
+  def string_subcatframe_withnodes(subcatframe)
+    return subcatframe.map { |gf, prep, fes, nodes| "#{gf} #{prep} " + nodes.map { |n| n.id.to_s }.join(",") }.sort.join(" ")
+  end
+
+end
+
+#####################################################################
+# main class
+#####################################################################
+
+class GfInduce
+
+  #########################################
+  # Initialization
+  #########################################
+
+  ###
+  # initialize everything to an empty hash,
+  # preparing for induce_from_sent.
+  # If you would like to start with induced GF already in,
+  # in order to use apply(), do GfInduce.from_file(filename)
+  #
+  # include_sem: if true, keep frame name and FE name
+  # as part of the subcat frame. if false, don't keep them
+  def initialize(interpreter_class, # SynInterpreter class
+		 include_sem = false)# boolean
+
+    @interpreter = interpreter_class
+    @gf_path_map = GfiGfPathMapping.new(interpreter_class)
+    @subcat_frames = GfiSubcatFrames.new(include_sem)
+  end
+
+  #########################################
+  # Pickling
+  #########################################
+
+  ###
+  # save this GfInduce object (as a pickle) to the given file
+  def to_file(filename) # string
+    begin
+      file = File.new(filename, "w")
+    rescue
+      $stderr.puts "GfInduce error: couldn't write to file #{filename}."
+      return
+    end
+
+    file.puts Marshal.dump(self)
+    file.close()
+  end
+
+  ###
+  # load a GfInduce object from the given file
+  # and return it.
+  # Returns nil if reading from the file failed.
+  def GfInduce.from_file(filename) # string
+    begin
+      file = File.new(filename)
+    rescue
+      $stderr.puts "GfInduce error: couldn't read from file #{filename}."
+      return nil
+    end
+
+    gfi_obj =  Marshal.load(file)
+    file.close()
+    return gfi_obj
+  end
+
+  #########################################
+  # Inducing mappings from training data
+  #########################################
+
+  ###
+  # induce path -> gf mapping from the given SalsaTigerSentence object
+  # 
+  # Assumption: sent contains semantic annotation: FrameNet frames
+  # and the FEs of the frames have information on grammatical function (gf)
+  # and phrase type (pt) of the phrase that the FE points to
+  # as attributes on FeNode objects (which represent <fe> elements in the
+  # underlying Salsa/Tiger XML representation)
+  def induce_from_sent(sent) # SalsaTigerSentence object
+
+    # induce GFs from each frame of the sentence
+    sent.each_frame { |frame|
+      unless frame.target
+        # frame without a target: 
+        # nothing I can do
+        next
+      end
+
+      # main target node, lemma
+      maintarget, targetlemma, targetpos = mainnode_and_lemma(frame.target.children())
+      if not(maintarget) or not(targetlemma)
+        # cannot count this one
+        next
+      end
+
+      # array of tuples [gfpt, prep, fe]
+      subcatframe = Array.new
+
+      # each FE (but not the target itself):
+      frame.each_child { |fe|
+        if fe.name == "target"
+          next
+        end
+
+        if not(fe.get_attribute("gf")) and not(fe.get_attribute("pt"))
+          # no GF or PT information: nothing to learn here
+          next
+        end
+
+        gfpt = "#{fe.get_attribute("gf")} #{fe.get_attribute("pt")}"
+
+        # compute path between main target and FE syn nodes,
+        # store mapping gfpt -> path in fngf_to_paths
+        fe.each_child { |syn_node|
+
+          # determine path,
+          path = @interpreter.path_between(maintarget, syn_node, true)
+
+          # store the mapping
+          @gf_path_map.store_mapping(gfpt, path, syn_node, targetlemma, targetpos)
+        
+          # preposition?
+          prep = @interpreter.preposition(syn_node)
+          if prep
+            prep.downcase!
+          end
+
+          # remember combination gfpt/prep/fe 
+          # as part of the subcat frame
+          subcatframe << [gfpt, prep, fe.name()]
+        } # each syn node that the FE points to
+      } # each FE of the frame
+
+      # store the subcat frame
+      @subcat_frames.store_subcatframe(subcatframe, frame.name(), targetlemma, targetpos)
+    } # each frame
+  end
+
+  ###
+  # finish up inducing:
+  #  reencode information in a fashion
+  #  that makes apply() faster
+  def compute_mapping()
+    @gf_path_map.finish_inducing()
+  end
+
+  #########################################
+  # Test output
+  #########################################
+
+  ###
+  def test_output()
+    @gf_path_map.test_output()
+    @subcat_frames.test_output()
+  end
+
+  #########################################
+  # Restricting induced mappings 
+  # to achieve better mappings
+  #########################################
+
+  ####
+  # restrict gf -> path mappings:
+  # exclude all paths that include an Up edge
+  def restrict_to_downpaths()
+    @gf_path_map.restrict_to_downpaths()
+  end
+
+  ####
+  # restrict gf -> path mappings:
+  # only keep paths up to length n
+  def restrict_pathlen(n)  # integer: maximum length to keep
+    @gf_path_map.restrict_pathlen(n)
+  end
+
+  ####
+  # restrict gf -> path mappings:
+  # remove GFs that are often incorrect
+  def remove_gfs(gf_list)
+    @gf_path_map.remove_gfs(gf_list)
+  end
+
+  #########################################
+  # Applying mappings to new data
+  #########################################
+
+
+
+  ###
+  # given a list of nodes (idea: they form a MWE together;
+  #  may of course be a single node),
+  # determine all subcat frames, i.e. all consistent sets of grammatical functions, 
+  # for the main node among the nodelist.
+  # For each subcat frame, potential FN frames and FE labels
+  # are returned as well
+  #
+  # strict: boolean. If true, return only those subcat frames that exactly match
+  #   all GFs listed in node_to_gf. If false, also return subcat frames that 
+  #   match a subset of the GFs mentioned in node_to_gf.
+  # 
+  #
+  # returns: list of tuples [frame(string), subcat frame, frequency(integer)], 
+  # where a subcat frame is an array of tuples 
+  # [gf (string), prep(string or nil), fe(string), synnodes(array:SynNode)]
+  def apply(nodelist, # array:SynNode
+            strict = false) # match: strict or subseteq?
+
+    mainnode, lemma, pos = mainnode_and_lemma(nodelist)
+    if not(mainnode) or not(lemma)
+      return []
+    end
+
+    unless @subcat_frames.lemma_known(lemma, pos)
+      # nothing known about the lemma
+      return []
+    end
+
+    # hash: SynNode -> tuple [GF(string), preposition(string), frequency(integer)]
+    node_to_gf = @gf_path_map.potential_gfs_of_node(mainnode, lemma, pos)
+
+#     $stderr.puts "HIER m:#{mainnode.to_s} l:#{lemma} p:{pos} "+ nodelist.map { |n| n.to_s}.join(" ")
+#     $stderr.puts "HIER2 #{@subcat_frames.lemma_known(lemma, pos)}"
+#     $stderr.puts "HIER3 #{node_to_gf.length()}"
+
+
+    return @subcat_frames.match(mainnode, lemma, pos, node_to_gf, strict)
+  end
+  
+
+  #########################################
+  #########################################
+  private
+
+  #########################################
+  # Main node, lemma, POS of given expression
+  #########################################
+
+  ###
+  # determine main node and its lemma
+  #
+  # returns: SynNode*string*string, main node, lemma, POS
+  def mainnode_and_lemma(nodelist)
+    mainnode = @interpreter.main_node_of_expr(nodelist)
+    unless mainnode
+      return [nil, nil, nil]
+    end
+
+    lemma = @interpreter.lemma_backoff(mainnode)
+    pos = @interpreter.category(mainnode)
+
+    # verb? then add the voice to the POS
+    if (voice = @interpreter.voice(mainnode))
+      pos = pos + "-" + voice
+    end    
+    return [mainnode, lemma, pos]
+  end
+
+end