mosesdecoder/scripts/training/mbr/mbr.cpp

#include <iostream>
#include <fstream>
#include <sstream>
#include <iomanip>
#include <vector>
#include <map>
#include <stdlib.h>
#include <math.h>
#include <algorithm>
#include <stdio.h>
#include <unistd.h>
#include <cstring>
#include <time.h>

using namespace std ;


/* Input :
   1. a sorted  n-best list, with duplicates filtered out in the following  format
   0 ||| amr moussa is currently on a visit to libya , tomorrow , sunday , to hold talks with regard to the in sudan . ||| 0 -4.94418 0 0 -2.16036 0 0 -81.4462 -106.593 -114.43 -105.55 -12.7873 -26.9057 -25.3715 -52.9336 7.99917 -24 ||| -4.58432

   2. a weight vector
   3. bleu order ( default = 4)
   4. scaling factor to weigh the weight vector (default = 1.0)

   Output :
   translations that minimise the Bayes Risk of the n-best list


*/

int TABLE_LINE_MAX_LENGTH = 5000;
vector<double> weights;
float SCALE = 1.0;
int BLEU_ORDER = 4;
int SMOOTH = 1;
int DEBUG = 0;
double min_interval = 1e-4;

#define SAFE_GETLINE(_IS, _LINE, _SIZE, _DELIM) {_IS.getline(_LINE, _SIZE, _DELIM); if(_IS.fail() && !_IS.bad() && !_IS.eof()) _IS.clear();}

typedef string WORD;
typedef unsigned int WORD_ID;


map<WORD, WORD_ID>  lookup;
vector< WORD > vocab;

class candidate_t
{
public:
  vector<WORD_ID> translation;
  vector<double> features;
  int translation_size;
} ;


void usage(void)
{
  fprintf(stderr,
          "usage: mbr -s SCALE -n BLEU_ORDER -w weights.txt -i nbest.txt");
}


char *strstrsep(char **stringp, const char *delim)
{
  char *match, *save;
  save = *stringp;
  if (*stringp == NULL)
    return NULL;
  match = strstr(*stringp, delim);
  if (match == NULL) {
    *stringp = NULL;
    return save;
  }
  *match = '\0';
  *stringp = match + strlen(delim);
  return save;
}



vector<string> tokenize( const char input[] )
{
  vector< string > token;
  bool betweenWords = true;
  int start;
  int i=0;
  for(; input[i] != '\0'; i++) {
    bool isSpace = (input[i] == ' ' || input[i] == '\t');

    if (!isSpace && betweenWords) {
      start = i;
      betweenWords = false;
    } else if (isSpace && !betweenWords) {
      token.push_back( string( input+start, i-start ) );
      betweenWords = true;
    }
  }
  if (!betweenWords)
    token.push_back( string( input+start, i-start+1 ) );
  return token;
}



WORD_ID storeIfNew( WORD word )
{
  if( lookup.find( word ) != lookup.end() )
    return lookup[ word ];

  WORD_ID id = vocab.size();
  vocab.push_back( word );
  lookup[ word ] = id;
  return id;
}

int count( string input, char delim )
{
  int count = 0;
  for ( int i = 0; i < input.size(); i++) {
    if ( input[i] == delim)
      count++;
  }
  return count;
}

double calculate_probability(const vector<double> & feats, const vector<double> & weights,double SCALE)
{

  if (feats.size() != weights.size())
    cerr << "ERROR : Number of features <> number of weights " << endl;

  double prob = 0;
  for ( int i = 0; i < feats.size(); i++) {
    prob += feats[i]*weights[i]*SCALE;
  }
  return exp(prob);
}

void extract_ngrams(const vector<WORD_ID>& sentence, map < vector < WORD_ID>, int >  & allngrams)
{
  vector< WORD_ID> ngram;
  for (int k = 0; k< BLEU_ORDER; k++) {
    for(int i =0; i < max((int)sentence.size()-k,0); i++) {
      for ( int j = i; j<= i+k; j++) {
        ngram.push_back(sentence[j]);
      }
      ++allngrams[ngram];
      ngram.clear();
    }
  }
}


double calculate_score(const vector<candidate_t*> & sents, int ref, int hyp,  vector < map < vector < WORD_ID>, int > > & ngram_stats )
{
  int comps_n = 2*BLEU_ORDER+1;
  int comps[comps_n];
  double logbleu = 0.0, brevity;

  int hyp_length = sents[hyp]->translation_size;

  for (int i =0; i<BLEU_ORDER; i++) {
    comps[2*i] = 0;
    comps[2*i+1] = max(hyp_length-i,0);
  }

  map< vector < WORD_ID > ,int > & hyp_ngrams = ngram_stats[hyp] ;
  map< vector < WORD_ID >, int > & ref_ngrams = ngram_stats[ref] ;

  for (map< vector< WORD_ID >, int >::iterator it = hyp_ngrams.begin();
       it != hyp_ngrams.end(); it++) {
    map< vector< WORD_ID >, int >::iterator ref_it = ref_ngrams.find(it->first);
    if(ref_it != ref_ngrams.end()) {
      comps[2* (it->first.size()-1)] += min(ref_it->second,it->second);
    }
  }
  comps[comps_n-1] = sents[ref]->translation_size;

  if (DEBUG) {
    for ( int i = 0; i < comps_n; i++)
      cerr << "Comp " << i << " : " << comps[i];
  }

  for (int i=0; i<BLEU_ORDER; i++) {
    if (comps[0] == 0)
      return 0.0;
    if ( i > 0 )
      logbleu += log(static_cast<double>(comps[2*i]+SMOOTH))-log(static_cast<double>(comps[2*i+1]+SMOOTH));
    else
      logbleu += log(static_cast<double>(comps[2*i]))-log(static_cast<double>(comps[2*i+1]));
  }
  logbleu /= BLEU_ORDER;
  brevity = 1.0-(double)comps[comps_n-1]/comps[1]; // comps[comps_n-1] is the ref length, comps[1] is the test length
  if (brevity < 0.0)
    logbleu += brevity;
  return exp(logbleu);
}

vector<double> read_weights(string fileName)
{
  ifstream inFile;
  inFile.open(fileName.c_str());
  istream *inFileP = &inFile;

  char line[TABLE_LINE_MAX_LENGTH];
  int i=0;
  vector<double> weights;

  while(true) {
    i++;
    SAFE_GETLINE((*inFileP), line, TABLE_LINE_MAX_LENGTH, '\n');
    if (inFileP->eof()) break;
    vector<string> token = tokenize(line);

    for (int j = 0; j < token.size(); j++) {
      weights.push_back(atof(token[j].c_str()));
    }
  }
  cerr << endl;
  return weights;
}

int find_pos_of_min_element(const vector<double>& vec)
{

  int min_pos = -1;
  double min_element = 10000;
  for ( int i = 0; i < vec.size(); i++) {
    if (vec[i] < min_element) {
      min_element = vec[i];
      min_pos = i;
    }
  }
  /*    cerr << "Min pos is : " << min_pos << endl;
      cerr << "Min mbr loss is : " << min_element << endl;*/
  return min_pos;
}

void process(int sent, const vector<candidate_t*> & sents)
{
//   cerr << "Sentence " << sent << " has " << sents.size() << " candidate translations" << endl;
  double marginal = 0;

  vector<double> joint_prob_vec;
  double joint_prob;
  vector< map < vector <WORD_ID>, int > > ngram_stats;

  for (int i = 0; i < sents.size(); i++) {
//         cerr << "Sents " << i << " has trans : " << sents[i]->translation << endl;
    //Calculate marginal and cache the posteriors
    joint_prob = calculate_probability(sents[i]->features,weights,SCALE);
    marginal += joint_prob;
    joint_prob_vec.push_back(joint_prob);
    //Cache ngram counts
    map < vector <WORD_ID>, int > counts;
    extract_ngrams(sents[i]->translation,counts);
    ngram_stats.push_back(counts);
  }
  //cerr << "Marginal is " << marginal;

  vector<double> mbr_loss;
  double bleu, weightedLoss;
  double weightedLossCumul = 0;
  double minMBRLoss = 1000000;
  int minMBRLossIdx = -1;

  /* Main MBR computation done here */
  for (int i = 0; i < sents.size(); i++) {
    weightedLossCumul = 0;
    for (int j = 0; j < sents.size(); j++) {
      if ( i != j) {
        bleu = calculate_score(sents, j, i,ngram_stats );
        weightedLoss = ( 1 - bleu) * ( joint_prob_vec[j]/marginal);
        weightedLossCumul += weightedLoss;
        if (weightedLossCumul > minMBRLoss)
          break;
      }
    }
    if (weightedLossCumul < minMBRLoss) {
      minMBRLoss = weightedLossCumul;
      minMBRLossIdx = i;
    }
  }
//    cerr << "Min pos is : " << minMBRLossIdx << endl;
//    cerr << "Min mbr loss is : " << minMBRLoss << endl;
  /* Find sentence that minimises Bayes Risk under 1- BLEU loss */
  vector<  WORD_ID > best_translation = sents[minMBRLossIdx]->translation;
  for (int i = 0; i < best_translation.size(); i++)
    cout << vocab[best_translation[i]] << " " ;
  cout << endl;
}


void read_nbest_data(string fileName)
{

  FILE * fp;
  fp = fopen (fileName.c_str() , "r");

  static char buf[10000];
  char *rest, *tok;
  int field;
  int sent_i, cur_sent;
  candidate_t *cand = NULL;
  vector<candidate_t*> testsents;

  cur_sent = -1;

  while (fgets(buf, sizeof(buf), fp) != NULL) {
    field = 0;
    rest = buf;
    while ((tok = strstrsep(&rest, "|||")) != NULL) {
      if (field == 0) {
        sent_i = strtol(tok, NULL, 10);
        cand = new candidate_t;
      } else if (field == 2) {
        vector<double> features;
        char * subtok;
        subtok = strtok (tok," ");

        while (subtok != NULL) {
          features.push_back(atof(subtok));
          subtok = strtok (NULL, " ");
        }
        cand->features = features;
      } else if (field == 1) {
        vector<string> trans_str = tokenize(tok);
        vector<WORD_ID> trans_int;
        for (int j=0; j<trans_str.size(); j++) {
          trans_int.push_back( storeIfNew( trans_str[j] ) );
        }
        cand->translation= trans_int;
        cand->translation_size = cand->translation.size();
      } else if (field == 3) {
        continue;
      } else {
        fprintf(stderr, "too many fields in n-best list line\n");
      }
      field++;
    }
    if (sent_i != cur_sent) {
      if (cur_sent != - 1) {
        process(cur_sent,testsents);
      }
      cur_sent = sent_i;
      testsents.clear();
    }
    testsents.push_back(cand);
  }
  process(cur_sent,testsents);
  cerr << endl;
}

int main(int argc, char **argv)
{

  time_t starttime = time(NULL);
  int c;

  string f_weight = "";
  string f_nbest = "";

  while ((c = getopt(argc, argv, "s:w:n:i:")) != -1) {
    switch (c) {
    case 's':
      SCALE = atof(optarg);
      break;
    case 'n':
      BLEU_ORDER = atoi(optarg);
      break;
    case 'w':
      f_weight =  optarg;
      break;
    case 'i':
      f_nbest = optarg;
      break;
    default:
      usage();
    }
  }

  argc -= optind;
  argv += optind;

  if (argc < 2) {
    usage();
  }


  weights = read_weights(f_weight);
  read_nbest_data(f_nbest);

  time_t endtime = time(NULL);
  cerr << "Processed data in" << (endtime-starttime) << " seconds\n";
}