translation/tools/mosesdecoder/mert/HwcmScorer.cpp

#include "HwcmScorer.h"

#include <fstream>

#include "ScoreStats.h"
#include "Util.h"

#include "util/tokenize_piece.hh"

// HWCM score (Liu and Gildea, 2005). Implements F1 instead of precision for better modelling of hypothesis length.
// assumes dependency trees on target side (generated by scripts/training/wrappers/conll2mosesxml.py ; use with option --brackets for reference).
// reads reference trees from separate file {REFERENCE_FILE}.trees to support mix of string-based and tree-based metrics.

using namespace std;

namespace MosesTuning
{


HwcmScorer::HwcmScorer(const string& config)
  : StatisticsBasedScorer("HWCM",config) {}

HwcmScorer::~HwcmScorer() {}

void HwcmScorer::setReferenceFiles(const vector<string>& referenceFiles)
{
  // For each line in the reference file, create a tree object
  if (referenceFiles.size() != 1) {
    throw runtime_error("HWCM only supports a single reference");
  }
  m_ref_trees.clear();
  m_ref_hwc.clear();
  ifstream in((referenceFiles[0] + ".trees").c_str());
  if (!in) {
    throw runtime_error("Unable to open " + referenceFiles[0] + ".trees");
  }
  string line;
  while (getline(in,line)) {
    line = this->preprocessSentence(line);
    TreePointer tree (boost::make_shared<InternalTree>(line));
    m_ref_trees.push_back(tree);
    vector<map<string, int> > hwc (kHwcmOrder);
    vector<string> history(kHwcmOrder);
    extractHeadWordChain(tree, history, hwc);
    m_ref_hwc.push_back(hwc);
    vector<int> totals(kHwcmOrder);
    for (size_t i = 0; i < kHwcmOrder; i++) {
      for (map<string, int>::const_iterator it = m_ref_hwc.back()[i].begin(); it != m_ref_hwc.back()[i].end(); it++) {
        totals[i] += it->second;
      }
    }
    m_ref_lengths.push_back(totals);
  }
  TRACE_ERR(endl);

}

void HwcmScorer::extractHeadWordChain(TreePointer tree, vector<string> & history, vector<map<string, int> > & hwc)
{

  if (tree->GetLength() > 0) {
    string head = getHead(tree);

    if (head.empty()) {
      for (std::vector<TreePointer>::const_iterator it = tree->GetChildren().begin(); it != tree->GetChildren().end(); ++it) {
        extractHeadWordChain(*it, history, hwc);
      }
    } else {
      vector<string> new_history(kHwcmOrder);
      new_history[0] = head;
      hwc[0][head]++;
      for (size_t hist_idx = 0; hist_idx < kHwcmOrder-1; hist_idx++) {
        if (!history[hist_idx].empty()) {
          string chain = history[hist_idx] + " " + head;
          hwc[hist_idx+1][chain]++;
          if (hist_idx+2 < kHwcmOrder) {
            new_history[hist_idx+1] = chain;
          }
        }
      }
      for (std::vector<TreePointer>::const_iterator it = tree->GetChildren().begin(); it != tree->GetChildren().end(); ++it) {
        extractHeadWordChain(*it, new_history, hwc);
      }
    }
  }
}

string HwcmScorer::getHead(TreePointer tree)
{
  // assumption (only true for dependency parse: each constituent has a preterminal label, and corresponding terminal is head)
  // if constituent has multiple preterminals, first one is picked; if it has no preterminals, empty string is returned
  for (std::vector<TreePointer>::const_iterator it = tree->GetChildren().begin(); it != tree->GetChildren().end(); ++it) {
    TreePointer child = *it;

    if (child->GetLength() == 1 && child->GetChildren()[0]->IsTerminal()) {
      return child->GetChildren()[0]->GetLabel();
    }
  }
  return "";

}

void HwcmScorer::prepareStats(size_t sid, const string& text, ScoreStats& entry)
{
  if (sid >= m_ref_trees.size()) {
    stringstream msg;
    msg << "Sentence id (" << sid << ") not found in reference set";
    throw runtime_error(msg.str());
  }

  string sentence = this->preprocessSentence(text);

  // if sentence has '|||', assume that tree is in second position (n-best-list);
  // otherwise, assume it is in first position (calling 'evaluate' with tree as reference)
  util::TokenIter<util::MultiCharacter> it(sentence, util::MultiCharacter("|||"));
  ++it;
  if (it) {
    sentence = it->as_string();
  }

  TreePointer tree (boost::make_shared<InternalTree>(sentence));
  vector<map<string, int> > hwc_test (kHwcmOrder);
  vector<string> history(kHwcmOrder);
  extractHeadWordChain(tree, history, hwc_test);

  ostringstream stats;
  for (size_t i = 0; i < kHwcmOrder; i++) {
    int correct = 0;
    int test_total = 0;
    for (map<string, int>::const_iterator it = hwc_test[i].begin(); it != hwc_test[i].end(); it++) {
      test_total += it->second;
      map<string, int>::const_iterator it2 = m_ref_hwc[sid][i].find(it->first);
      if (it2 != m_ref_hwc[sid][i].end()) {
        correct += std::min(it->second, it2->second);
      }
    }
    stats << correct << " " << test_total << " " << m_ref_lengths[sid][i] << " " ;
  }

  string stats_str = stats.str();
  entry.set(stats_str);
}

float HwcmScorer::calculateScore(const vector<ScoreStatsType>& comps) const
{
  float precision = 0;
  float recall = 0;
  for (size_t i = 0; i < kHwcmOrder; i++) {
    float matches = comps[i*3];
    float test_total = comps[1+(i*3)];
    float ref_total = comps[2+(i*3)];
    if (test_total > 0) {
      precision += matches/test_total;
    }
    if (ref_total > 0) {
      recall += matches/ref_total;
    }
  }

  precision /= (float)kHwcmOrder;
  recall /= (float)kHwcmOrder;
  return (2*precision*recall)/(precision+recall); // f1-score
}

}