diff options
Diffstat (limited to 'training/dtrain/score.h')
| -rw-r--r-- | training/dtrain/score.h | 200 |
1 files changed, 92 insertions, 108 deletions
diff --git a/training/dtrain/score.h b/training/dtrain/score.h index 1cdd3fa9..d51aef82 100644 --- a/training/dtrain/score.h +++ b/training/dtrain/score.h @@ -6,20 +6,19 @@ namespace dtrain { - struct NgramCounts { - unsigned N_; - map<unsigned, score_t> clipped_; - map<unsigned, score_t> sum_; + size_t N_; + map<size_t, weight_t> clipped_; + map<size_t, weight_t> sum_; - NgramCounts(const unsigned N) : N_(N) { Zero(); } + NgramCounts(const size_t N) : N_(N) { Zero(); } inline void operator+=(const NgramCounts& rhs) { if (rhs.N_ > N_) Resize(rhs.N_); - for (unsigned i = 0; i < N_; i++) { + for (size_t i = 0; i < N_; i++) { this->clipped_[i] += rhs.clipped_.find(i)->second; this->sum_[i] += rhs.sum_.find(i)->second; } @@ -30,20 +29,21 @@ struct NgramCounts { NgramCounts result = *this; result += other; + return result; } inline void - operator*=(const score_t rhs) + operator*=(const weight_t rhs) { - for (unsigned i = 0; i < N_; i++) { + for (size_t i = 0; i < N_; i++) { this->clipped_[i] *= rhs; this->sum_[i] *= rhs; } } inline void - Add(const unsigned count, const unsigned ref_count, const unsigned i) + Add(const size_t count, const size_t ref_count, const size_t i) { assert(i < N_); if (count > ref_count) { @@ -57,40 +57,31 @@ struct NgramCounts inline void Zero() { - for (unsigned i = 0; i < N_; i++) { + for (size_t i = 0; i < N_; i++) { clipped_[i] = 0.; sum_[i] = 0.; } } inline void - One() - { - for (unsigned i = 0; i < N_; i++) { - clipped_[i] = 1.; - sum_[i] = 1.; - } - } - - inline void - Print() + Print(ostream& os=cerr) { - for (unsigned i = 0; i < N_; i++) { - cout << i+1 << "grams (clipped):\t" << clipped_[i] << endl; - cout << i+1 << "grams:\t\t\t" << sum_[i] << endl; + for (size_t i = 0; i < N_; i++) { + os << i+1 << "grams (clipped):\t" << clipped_[i] << endl; + os << i+1 << "grams:\t\t\t" << sum_[i] << endl; } } - inline void Resize(unsigned N) + inline void Resize(size_t N) { if (N == N_) return; else if (N > N_) { - for (unsigned i = N_; i < N; i++) { + for (size_t i = N_; i < N; i++) { clipped_[i] = 0.; sum_[i] = 0.; } } else { // N < N_ - for (unsigned i = N_-1; i > N-1; i--) { + for (size_t i = N_-1; i > N-1; i--) { clipped_.erase(i); sum_.erase(i); } @@ -99,123 +90,116 @@ struct NgramCounts } }; -typedef map<vector<WordID>, unsigned> Ngrams; +typedef map<vector<WordID>, size_t> Ngrams; inline Ngrams -make_ngrams(const vector<WordID>& s, const unsigned N) +MakeNgrams(const vector<WordID>& s, const size_t N) { Ngrams ngrams; vector<WordID> ng; for (size_t i = 0; i < s.size(); i++) { ng.clear(); - for (unsigned j = i; j < min(i+N, s.size()); j++) { + for (size_t j = i; j < min(i+N, s.size()); j++) { ng.push_back(s[j]); ngrams[ng]++; } } + return ngrams; } inline NgramCounts -make_ngram_counts(const vector<WordID>& hyp, const vector<WordID>& ref, const unsigned N) +MakeNgramCounts(const vector<WordID>& hyp, + const vector<Ngrams>& ref, + const size_t N) { - Ngrams hyp_ngrams = make_ngrams(hyp, N); - Ngrams ref_ngrams = make_ngrams(ref, N); + Ngrams hyp_ngrams = MakeNgrams(hyp, N); NgramCounts counts(N); - Ngrams::iterator it; - Ngrams::iterator ti; + Ngrams::iterator it, ti; for (it = hyp_ngrams.begin(); it != hyp_ngrams.end(); it++) { - ti = ref_ngrams.find(it->first); - if (ti != ref_ngrams.end()) { - counts.Add(it->second, ti->second, it->first.size() - 1); - } else { - counts.Add(it->second, 0, it->first.size() - 1); + size_t max_ref_count = 0; + for (auto r: ref) { + ti = r.find(it->first); + if (ti != r.end()) + max_ref_count = max(max_ref_count, ti->second); } + counts.Add(it->second, min(it->second, max_ref_count), it->first.size()-1); } + return counts; } -struct BleuScorer : public LocalScorer +/* + * per-sentence BLEU + * as in "Optimizing for Sentence-Level BLEU+1 + * Yields Short Translations" + * (Nakov et al. '12) + * + * [simply add 1 to reference length for calculation of BP] + * + */ +struct PerSentenceBleuScorer { - score_t Bleu(NgramCounts& counts, const unsigned hyp_len, const unsigned ref_len); - score_t Score(const vector<WordID>& hyp, const vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); - void Reset() {} -}; - -struct StupidBleuScorer : public LocalScorer -{ - score_t Score(const vector<WordID>& hyp, const vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); - void Reset() {} -}; - -struct FixedStupidBleuScorer : public LocalScorer -{ - score_t Score(const vector<WordID>& hyp, const vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); - void Reset() {} -}; - -struct SmoothBleuScorer : public LocalScorer -{ - score_t Score(const vector<WordID>& hyp, const vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); - void Reset() {} -}; - -struct SumBleuScorer : public LocalScorer -{ - score_t Score(const vector<WordID>& hyp, const vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); - void Reset() {} -}; + const size_t N_; + vector<weight_t> w_; -struct SumExpBleuScorer : public LocalScorer -{ - score_t Score(const vector<WordID>& hyp, const vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); - void Reset() {} -}; - -struct SumWhateverBleuScorer : public LocalScorer -{ - score_t Score(const vector<WordID>& hyp, const vector<WordID>& ref, const unsigned /*rank*/, const unsigned /*src_len*/); - void Reset() {}; -}; - -struct ApproxBleuScorer : public BleuScorer -{ - NgramCounts glob_onebest_counts_; - unsigned glob_hyp_len_, glob_ref_len_, glob_src_len_; - score_t discount_; - - ApproxBleuScorer(unsigned N, score_t d) : glob_onebest_counts_(NgramCounts(N)), discount_(d) + PerSentenceBleuScorer(size_t n) : N_(n) { - glob_hyp_len_ = glob_ref_len_ = glob_src_len_ = 0; + for (size_t i = 1; i <= N_; i++) + w_.push_back(1.0/N_); } - inline void Reset() { - glob_onebest_counts_.Zero(); - glob_hyp_len_ = glob_ref_len_ = glob_src_len_ = 0.; - } - - score_t Score(const vector<WordID>& hyp, const vector<WordID>& ref, const unsigned rank, const unsigned src_len); -}; - -struct LinearBleuScorer : public BleuScorer -{ - unsigned onebest_len_; - NgramCounts onebest_counts_; - - LinearBleuScorer(unsigned N) : onebest_len_(1), onebest_counts_(N) + inline weight_t + BrevityPenalty(const size_t hl, const size_t rl) { - onebest_counts_.One(); + if (hl > rl) + return 1; + + return exp(1 - (weight_t)rl/hl); } - score_t Score(const vector<WordID>& hyp, const vector<WordID>& ref, const unsigned rank, const unsigned /*src_len*/); + weight_t + Score(const vector<WordID>& hyp, + const vector<Ngrams>& ref_ngs, + const vector<size_t>& ref_ls) + { + size_t hl = hyp.size(), rl = 0; + if (hl == 0) return 0.; + // best match reference length + if (ref_ls.size() == 1) { + rl = ref_ls.front(); + } else { + size_t i = 0, best_idx = 0; + size_t best = numeric_limits<size_t>::max(); + for (auto l: ref_ls) { + size_t d = abs(hl-l); + if (d < best) { + best_idx = i; + best = d; + } + i += 1; + } + rl = ref_ls[best_idx]; + } + if (rl == 0) return 0.; + NgramCounts counts = MakeNgramCounts(hyp, ref_ngs, N_); + size_t M = N_; + vector<weight_t> v = w_; + if (rl < N_) { + M = rl; + for (size_t i = 0; i < M; i++) v[i] = 1/((weight_t)M); + } + weight_t sum = 0, add = 0; + for (size_t i = 0; i < M; i++) { + if (i == 0 && (counts.sum_[i] == 0 || counts.clipped_[i] == 0)) return 0.; + if (i == 1) add = 1; + sum += v[i] * log(((weight_t)counts.clipped_[i] + add)/((counts.sum_[i] + add))); + } - inline void Reset() { - onebest_len_ = 1; - onebest_counts_.One(); + return BrevityPenalty(hl, rl+1) * exp(sum); } }; - } // namespace #endif |