zipf 1.2.1 → 1.2.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA1:
-  metadata.gz: 4a5374ff741398bdd58da174d27a130599c0573d
-  data.tar.gz: 4df6f82132a2a8c79d3af0608fa91de6515d2527
+  metadata.gz: 5dc46e50a6d5f63fb3b1db68dce9cb35a8342172
+  data.tar.gz: d849ab3031db7b9b65f77e16d746792dcdaf4827
 SHA512:
-  metadata.gz: 2a4ee64270b0af54115f2449774c6e632e9dc405f34169473318cb9c8fdccb52dfa470a5d7458f46bbf55e093086092354c38cbf05c4262a58497fd522540606
-  data.tar.gz: b8d1c18f2bda9ad022067ca2296aacfe71f0052ac8f0f54bf58178dae251f44ff07846f51e21f478e4ff6ae50df5a878591013c1bbdee2c80af2b7e9693cd7f8
+  metadata.gz: 12d97d1af9aa19d63ed02b2c17c0b3506fecf9ed0b0c9cbea4f4e233ff87e4734c3a8a4c0f1754a9d09cdb48072d249e1479d89fae5b5175eb242d43e3b24aa7
+  data.tar.gz: 9e179f4165835a25984547f1f6a98166dd36747ec966f7adbc8bc5cbe561627040fc6ab7b03ff61f551b3fac0293d6c49ac117ef9d37b969cfd33fdb271482ea

data/lib/zipf/bleu.rb

@@ -62,19 +62,34 @@ class BLEU::Ngrams
   end
 end
 
-def BLEU::get_counts hypothesis, reference, n, times=1
+def BLEU::best_match_length hypothesis, references
+  hyp_len = hypothesis.strip.split.size 
+  ref_lens = references.map { |r| r.strip.split.size }
+  min = Integer::MAX
+  min_idx = -1
+  ref_lens.each_with_index { |l,i|
+    min_idx = i if (hyp_len-l).abs < min
+  }
+  return hyp_len, ref_lens[min_idx]
+end
+
+def BLEU::get_counts hypothesis, references, n, times=1
   p = NgramCounts.new n
-  r = Ngrams.new
-  ngrams(reference, n) { |ng| r.add ng }
+  r = []
+  references.each { |reference|
+    r << Ngrams.new
+    ngrams(reference, n) { |ng| r.last.add ng }
+  }
   h = Ngrams.new
   ngrams(hypothesis, n) { |ng| h.add ng }
   h.each { |ng,count|
     sz = ng.size-1
     p.sum[sz] += count * times
-    p.clipped[sz] += [r.get_count(ng), count].min * times
+    p.clipped[sz] += [r.map { |i| i.get_count(ng)}.max, count].min * times
   }
-  p.ref_len = tokenize(reference.strip).size * times
-  p.hyp_len = tokenize(hypothesis.strip).size * times
+  p.hyp_len, p.ref_len = best_match_length hypothesis, references
+  p.hyp_len *= times
+  p.ref_len *= times
   return p
 end
 
@@ -82,45 +97,65 @@ def BLEU::brevity_penalty c, r, smooth=0.0
   return [0.0, 1.0-((r+smooth)/c)].min
 end
 
-def BLEU::bleu counts, n, debug=false
+def BLEU::bleu_ counts, n, debug=false
   corpus_stats = NgramCounts.new n
   counts.each { |i| corpus_stats.plus_eq i }
   logbleu = 0.0
   0.upto(n-1) { |m|
-    STDERR.write "#{m+1} #{corpus_stats.clipped[m]} / #{corpus_stats.sum[m]}\n" if debug
+    STDERR.write "#{m+1} #{corpus_stats.clipped[m]} / #{corpus_stats.sum[m]} = #{(corpus_stats.clipped[m]/corpus_stats.sum[m]).round 2}\n" if debug
     return 0.0 if corpus_stats.clipped[m] == 0 or corpus_stats.sum == 0
     logbleu += Math.log(corpus_stats.clipped[m]) - Math.log(corpus_stats.sum[m])
   }
   logbleu /= n
-  if debug
-    STDERR.write "BP #{brevity_penalty(corpus_stats.hyp_len, corpus_stats.ref_len)}\n"
-    STDERR.write "sum #{Math.exp(sum)}\n"
-  end
+  STDERR.write "BP #{brevity_penalty(corpus_stats.hyp_len, corpus_stats.ref_len).round 2}\n" if debug
   logbleu += brevity_penalty corpus_stats.hyp_len, corpus_stats.ref_len
   return Math.exp logbleu
 end
 
-def BLEU::hbleu counts, n, debug=false
+def BLEU::bleu hyp_file, ref_file, n, debug=false
+  hypotheses = ReadFile.readlines_strip(hyp_file)
+  references = ReadFile.readlines_strip(ref_file).map { |l|
+    splitpipe(l,3)
+  }
+  counts = []
+  hypotheses.each_with_index { |h,i|
+    counts << BLEU::get_counts(h, references[i], 4)
+  }
+  bleu_ counts, n, debug
+end
+
+def BLEU::hbleu_ counts, n, debug=false
   (100*bleu(counts, n, debug)).round(3)
 end
 
-def BLEU::per_sentence_bleu hypothesis, reference, n=4, smooth=0.0
-  h_ng = {}; r_ng = {}
-  (1).upto(n) { |i| h_ng[i] = []; r_ng[i] = [] }
+def BLEU::hbleu hypotheses, references, n, debug=false
+end
+
+def BLEU::per_sentence_bleu hypothesis, references, n=4, smooth=0.0
+  h_ng = {}; r_ng = []
+  num_ref = references.size
+  num_ref.times { r_ng << {} }
+  (1).upto(n) { |i| h_ng[i] = []; num_ref.times { |j| r_ng[j][i] = [] } }
   ngrams(hypothesis, n) { |i| h_ng[i.size] << i }
-  ngrams(reference, n) { |i| r_ng[i.size] << i }
-  m = [n, reference.split.size].min
+  references.each_with_index { |reference,j|
+    ngrams(reference, n) { |i| r_ng[j][i.size] << i }
+  }
+  m = [n, references.map { |i| i.split.size }.max].min
   add = 0.0
   logbleu = 0.0
   (1).upto(m) { |i|
     counts_clipped = 0
     counts_sum = h_ng[i].size
-    h_ng[i].uniq.each { |j| counts_clipped += r_ng[i].count(j) }
+    h_ng[i].uniq.each { |j|
+      max_count = [h_ng[i].count(j), r_ng.map { |r| r[i].count(j) }.max].min
+      counts_clipped += max_count 
+    }
     add = 1.0 if i >= 2
     logbleu += Math.log(counts_clipped+add) - Math.log(counts_sum+add);
   }
   logbleu /= m
-  logbleu += brevity_penalty hypothesis.strip.split.size, reference.strip.split.size, smooth
+  hyp_len, best_ref_len = BLEU::best_match_length hypothesis, references
+  logbleu += brevity_penalty hyp_len, best_ref_len, smooth
   return Math.exp logbleu
 end
 

data/lib/zipf/misc.rb

@@ -111,3 +111,11 @@ def read_config fn
   return cfg
 end
 
+# https://gist.github.com/pithyless/9738125
+class Integer
+  N_BYTES = [42].pack('i').size
+  N_BITS = N_BYTES * 16
+  MAX = 2 ** (N_BITS - 2) - 1
+  MIN = -MAX - 1
+end
+

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: zipf
 version: !ruby/object:Gem::Version
-  version: 1.2.1
+  version: 1.2.2
 platform: ruby
 authors:
 - Patrick Simianer
 autorequire: 
 bindir: bin
 cert_chain: []
-date: 2015-01-24 00:00:00.000000000 Z
+date: 2015-01-25 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: json