/** * @file   ngram_decode.c *  * <JA> * @brief  N-gram澄唯に答づく肌帽胳徒盧∈媽2パス∷  * * Julius のN-gramを脫いたスタックデコ〖ディング(媽2パス)において· * 肌に儡魯しうる帽胳の礁圭を瘋年する.  *  * 涂えられた鷗倡傅簿棱の幌眉フレ〖ムを徒盧し·帽胳トレリス懼で * その徒盧フレ〖ム件收に姜眉が賂哼する帽胳の礁圭を· * そのN-gram叫附澄唯とともに手す.  * * Julius では ngram_firstwords(), ngram_nextwords(), ngram_acceptable() が * それぞれ媽2パスのメイン簇眶 wchmm_fbs() から鈣び叫される. なお· * Julian ではこれらの簇眶の洛わりに dfa_decode.c の簇眶が脫いられる.  * </JA> *  * <EN> * @brief N-gram based word prediction for the 2nd pass. * * These functions returns next word candidates in the 2nd recognition * pass of Julius, i.e. N-gram based stack decoding. * * Given a partial sentence hypothesis, it first estimate the beginning frame * of the hypothesis based on the word trellis.  Then the words in the word * trellis around the estimated frame are extracted from the word trellis. * They will be returned with their N-gram probabilities. * * In Julius, ngram_firstwords(), ngram_nextwords() and ngram_acceptable() * are called from main search function wchmm_fbs().  In Julian, * corresponding functions in dfa_decode.c will be used instead. * </EN> *  * @author Akinobu Lee * @date   Fri Jul  8 14:57:51 2005 * * $Revision: 1.3 $ *  *//* * Copyright (c) 1991-2007 Kawahara Lab., Kyoto University * Copyright (c) 2000-2005 Shikano Lab., Nara Institute of Science and Technology * Copyright (c) 2005-2007 Julius project team, Nagoya Institute of Technology * All rights reserved */   #include <julius/julius.h>/**  * <JA> * 肌帽胳鉻輸ソ〖ト脫 qsort コ〖ルバック簇眶.  *  * @param a [in] 妥燎1 * @param b [in] 妥燎2 *  * @return aの帽胳ID > bの帽胳ID なら1, 嫡なら -1, 票じなら 0 を手す.  * </JA> * <EN> * qsort callback function to sort next word candidates by their word ID. *  * @param a [in] element 1 * @param b [in] element 2 *  * @return 1 if word id of a > that of b, -1 if negative, 0 if equal. * </EN> */static intcompare_nw(NEXTWORD **a, NEXTWORD **b){  if ((*a)->id > (*b)->id) return 1;  if ((*a)->id < (*b)->id) return -1;  return 0;}/**  * <JA> * 肌帽胳鉻輸リスト柒から帽胳を浮瑚する.  *  * @param nw [in] 肌帽胳鉻輸リスト * @param w [in] 浮瑚する帽胳のID * @param num [in] 肌帽胳鉻輸リストの墓さ *  * @return 斧つかった眷圭その肌帽胳鉻輸菇隴攣へのポインタ·斧つからなければ * NULL を手す.  * </JA> * <EN> * Find a word from list of next word candidates. *  * @param nw [in] list of next word candidates * @param w [in] word id to search for * @param num [in] length of @a nw *  * @return the pointer to the NEXTWORD data if found, or NULL if not found. * </EN> *//* find next word candiate whose id 'w' */static NEXTWORD *search_nw(NEXTWORD **nw, WORD_ID w, int num){  int left,right,mid;  NEXTWORD *tmp;  if (num == 0) return NULL;  left = 0;  right = num - 1;  while (left < right) {    mid = (left + right) / 2;    if ((nw[mid])->id < w) {      left = mid + 1;    } else {      right = mid;    }  }  tmp = nw[left];  if (tmp->id == w) {    return tmp;  } else {    return NULL;  }}/**  * <EN> * Compute backward N-gram score from forward N-gram. * </EN> * <JA> * 稿羹きの N-gram スコアを漣羹き N-gram から換叫する.  * </JA> *  * @param ngram [in] N-gram data structure * @param w [in] word sequence * @param wlen [in] length of @a w *  * @return the backward probability of the word w[0]. *  */static LOGPROBngram_forw2back(NGRAM_INFO *ngram, WORD_ID *w, int wlen){  int i;  LOGPROB p1, p2;  p1 = 0.0;  for(i = 1; i < ngram->n; i++) {    if (i >= wlen) break;    p1 += ngram_prob(ngram, i, &(w[1]));  }  p2 = 0.0;  for(i = 0; i < ngram->n; i++) {    if (i >= wlen) break;    p2 += ngram_prob(ngram, i+1, w);  }  return(p2 - p1);}/**  * <JA> * @brief  帽胳トレリスから肌帽胳鉻輸を藐叫する.  * * 帽胳トレリス懼の回年したフレ〖ム懼に姜眉が賂哼するトレリス帽胳 * のリストを藐叫し·それらの肌帽胳としての N-gram 儡魯澄唯を紛換する.  * そのリストを肌帽胳攫鼠菇隴攣に納裁して手す.  *  * @param r [in] 千急借妄インスタンス * @param nw [i/o] 肌帽胳鉻輸リスト∈藐叫馮蔡は @a oldnum 笆慣に納裁される∷ * @param oldnum [in] @a nw にすでに呈羌されている肌帽胳の眶 * @param hypo [in] 鷗倡傅の矢簿棱 * @param t [in] 回年フレ〖ム *  * @return 藐叫リストを納裁したあとの @a nw に崔まれる肌帽胳の另眶.  * </JA> * <EN> * @brief  Extract next word candidates from word trellis. * * This function extracts the list of trellis words whose word end * has survived in the word trellis at the specified frame. * The N-gram probabilities of them are then computed and added to * the current next word candidates data. *  * @param r [in] recognition process instance * @param nw [in] list of next word candidates (new words will be appended at @a oldnum) * @param oldnum [in] number of words already stored in @a nw * @param hypo [in] the source sentence hypothesis * @param t [in] specified frame *  * @return the total number of words currently stored in the @a nw. * </EN> */static intpick_backtrellis_words(RecogProcess *r, NEXTWORD **nw, int oldnum, NODE *hypo, short t){  int i;  WORD_ID w;  LOGPROB rawscore;#ifdef WPAIR  int w_old = WORD_INVALID;#endif  int num;  WORD_ID cnword[MAX_N];	///< Last two non-transparent words  WORD_ID cnwordrev[MAX_N];	///< Last two non-transparent words  int cnnum;		///< Num of found non-transparent words (<=2)  int last_trans;		///< Num of skipped transparent words  BACKTRELLIS *bt;  WORD_INFO *winfo;  NGRAM_INFO *ngram;  LOGPROB lm_weight2, lm_penalty2, lm_penalty_trans;  num = oldnum;  bt = r->backtrellis;  winfo = r->lm->winfo;  ngram = r->lm->ngram;  lm_weight2 = r->config->lmp.lm_weight2;  lm_penalty2 = r->config->lmp.lm_penalty2;  lm_penalty_trans = r->config->lmp.lm_penalty_trans;  /* set word contexts to cnword[] from 1 considering transparent words */  if (ngram) {    cnnum = 0;    last_trans = 0;    for(i=hypo->seqnum-1;i>=0;i--) {      if (! winfo->is_transparent[hypo->seq[i]]) {	cnword[cnnum+1] = hypo->seq[i];	cnnum++;	if (cnnum >= ngram->n - 1) break;      } else {	last_trans++;      }    }    if (ngram->dir == DIR_RL) {      for(i=0;i<cnnum;i++) {	cnwordrev[cnnum-1-i] = cnword[i+1];      }    }    /* use ngram id */    if (ngram->dir == DIR_RL) {      for(i=0;i<cnnum;i++) cnwordrev[i] = winfo->wton[cnwordrev[i]];    } else {      for(i=0;i<cnnum;i++) cnword[i+1] = winfo->wton[cnword[i+1]];    }  }  /* lookup survived words in backtrellis on time frame 't' */  for (i=0;i<bt->num[t];i++) {    w = (bt->rw[t][i])->wid;#ifdef WORD_GRAPH    /* only words on the word graphs are expanded */    if (!(bt->rw[t][i])->within_wordgraph) continue;#endif /* not WORD_GRAPH */#ifdef WPAIR    /* some word have same word ID with different previous word, so       only one will be opened (best word will be selected later       by next_word() */    if (w == w_old) continue;	/* backtrellis is sorted by word ID */    else w_old = w;#endif /* WPAIR */    /* skip if already exist */    if (search_nw(nw, w, oldnum) != NULL) continue;    /* compute LM probability of the word */    if (ngram) {      /* compute N-gram probability */      if (ngram->dir == DIR_RL) {	/* just compute N-gram prob of the word candidate */	cnwordrev[cnnum] = winfo->wton[w];	rawscore = ngram_prob(ngram, cnnum + 1, cnwordrev);      } else {	cnword[0] = winfo->wton[w];	rawscore = ngram_forw2back(ngram, cnword, cnnum + 1);      }#ifdef CLASS_NGRAM      rawscore += winfo->cprob[w];#endif    }    if (r->lmvar == LM_NGRAM_USER) {      /* call user-defined function */      /* be careful that the word context is ordered in backward direction */      rawscore = (*(r->lm->lmfunc.lmprob))(winfo, hypo->seq, hypo->seqnum, w, rawscore);    }    nw[num]->tre   = bt->rw[t][i];    nw[num]->id    = w;    nw[num]->lscore = rawscore * lm_weight2 + lm_penalty2;    if (winfo->is_transparent[w]) {      /*nw[num]->lscore -= (LOGPROB)last_trans * TRANS_RENZOKU_PENALTY;*/      if (winfo->is_transparent[hypo->seq[hypo->seqnum-1]]) {	nw[num]->lscore += lm_penalty_trans;      }    }        /* j_printf("%d: %s added\n", num, winfo->wname[nw[num]->id]); */    num++;  }  return num;}/**  * <JA> * @brief  帽胳トレリスから肌帽胳礁圭を瘋年する.  * * 回年フレ〖ムの漣稿 lookup_range 尸に姜眉があるトレリス懼の帽胳を礁め· * 肌帽胳菇隴攣を菇蜜する. 票じ帽胳が懼淡の認跋柒に剩眶ある眷圭· * 回年フレ〖ムにもっとも奪いトレリス懼の帽胳が聯買される.  *  * @param r [in] 千急借妄インスタンス * @param nw [out] 肌帽胳礁圭を呈羌する菇隴攣へのポインタ * @param hypo [in] 鷗倡傅の嬸尸矢簿棱 * @param tm [in] 帽胳を玫す面看となる回年フレ〖ム * @param t_end [in] 帽胳を玫すフレ〖ムの寶眉 *  * @return @a nw に呈羌された肌帽胳鉻輸の眶を手す.  * </JA> * <EN> * @brief  Determine next word candidates from the word trellis.