module BLEU class BLEU::NgramCounts attr_accessor :sum, :clipped, :ref_len, :hyp_len, :n def initialize(n) @n = 0 @sum = [] @clipped = [] @ref_len = 0.0 @hyp_len = 0.0 grow(n) end def grow(n) (n-@n).times { @sum << 0.0 @clipped << 0.0 } @n = n end def plus_eq(other) if other.n > @n then grow(other.n) end 0.upto(other.n-1) { |m| @sum[m] += other.sum[m] @clipped[m] += other.clipped[m] } @ref_len += other.ref_len @hyp_len += other.hyp_len end def to_s return "n=#{n} sum=#{sum} clipped=#{clipped} ref_len=#{ref_len} hyp_len=#{hyp_len}" end end class BLEU::Ngrams def initialize @h_ = {} @h_.default = 0 end def add(k) if k.class == Array then k = k.join ' ' end @h_[k] += 1 end def get_count(k) if k.class == Array then k = k.join ' ' end return @h_[k] end def each @h_.each_pair { |k,v| yield k.split, v } end def to_s @h_.to_s end end def BLEU::get_counts hypothesis, reference, n, times=1 p = NgramCounts.new n r = Ngrams.new ngrams(reference, n) { |ng| r.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.ref_len = tokenize(reference.strip).size * times p.hyp_len = tokenize(hypothesis.strip).size * times return p end def BLEU::brevity_penalty c, r, hack=0.0 return 1.0 if c>r return Math.exp 1.0-((r+hack)/c) end def BLEU::bleu counts, n, debug=false corpus_stats = NgramCounts.new n counts.each { |i| corpus_stats.plus_eq i } sum = 0.0 w = 1.0/n 0.upto(n-1) { |m| STDERR.write "#{m+1} #{corpus_stats.clipped[m]} / #{corpus_stats.sum[m]}\n" if debug return 0.0 if corpus_stats.clipped[m] == 0 or corpus_stats.sum == 0 sum += w * Math.log(corpus_stats.clipped[m] / corpus_stats.sum[m]) } if debug STDERR.write "BP #{brevity_penalty(corpus_stats.hyp_len, corpus_stats.ref_len)}\n" STDERR.write "sum #{Math.exp(sum)}\n" end return brevity_penalty(corpus_stats.hyp_len, corpus_stats.ref_len) * Math.exp(sum) 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, hack=0.0 h_ng = {}; r_ng = {} (1).upto(n) {|i| h_ng[i] = []; r_ng[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 weight = 1.0/m add = 0.0 sum = 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) } add = 1.0 if i >= 2 sum += weight * Math.log((counts_clipped + add)/(counts_sum + add)); } return brevity_penalty(hypothesis.strip.split.size, reference.strip.split.size) * Math.exp(sum) end end # module