libjulius/src/multi-gram.c

説明を見る。

/*
 * Copyright (c) 1991-2011 Kawahara Lab., Kyoto University
 * Copyright (c) 2000-2005 Shikano Lab., Nara Institute of Science and Technology
 * Copyright (c) 2005-2011 Julius project team, Nagoya Institute of Technology
 * All rights reserved
 */


#include <julius/julius.h>

#define MDEBUG

static boolean
multigram_rebuild_wchmm(RecogProcess *r)
{
  boolean ret;

  /* re-build wchmm */
  if (r->wchmm != NULL) {
    wchmm_free(r->wchmm);
  }
  r->wchmm = wchmm_new();
  r->wchmm->lmtype = r->lmtype;
  r->wchmm->lmvar  = r->lmvar;
  r->wchmm->ccd_flag = r->ccd_flag;
  r->wchmm->category_tree = TRUE;
  r->wchmm->hmmwrk = &(r->am->hmmwrk);
  /* assign models */
  r->wchmm->dfa = r->lm->dfa;
  r->wchmm->winfo = r->lm->winfo;
  r->wchmm->hmminfo = r->am->hmminfo;
  if (r->wchmm->category_tree) {
    if (r->config->pass1.old_tree_function_flag) {
      ret = build_wchmm(r->wchmm, r->lm->config);
    } else {
      ret = build_wchmm2(r->wchmm, r->lm->config);
    }
  } else {
    ret = build_wchmm2(r->wchmm, r->lm->config);
  }

  /* 起動時 -check でチェックモードへ */
  if (r->config->sw.wchmm_check_flag) {
    wchmm_check_interactive(r->wchmm);
  }
  
  if (ret == FALSE) {
    jlog("ERROR: multi-gram: failed to build (global) lexicon tree for recognition\n");
    return FALSE;
  }
  
  /* guess beam width from models, when not specified */
  r->trellis_beam_width = set_beam_width(r->wchmm, r->config->pass1.specified_trellis_beam_width);
  switch(r->config->pass1.specified_trellis_beam_width) {
  case 0:
    jlog("STAT: multi-gram: beam width set to %d (full) by lexicon change\n", r->trellis_beam_width);
    break;
  case -1:
    jlog("STAT: multi-gram: beam width set to %d (guess) by lexicon change\n", r->trellis_beam_width);
  }

  /* re-allocate factoring cache for the tree lexicon*/
  /* for n-gram only?? */
  //max_successor_cache_free(recog->wchmm);
  //max_successor_cache_init(recog->wchmm);

  /* finished! */

  return TRUE;
}

boolean
multigram_build(RecogProcess *r)
{
  if (r->lm->winfo != NULL) {
    /* re-build tree lexicon for recognition process */
    if (multigram_rebuild_wchmm(r) == FALSE) {
      jlog("ERROR: multi-gram: failed to re-build tree lexicon\n");
      return FALSE;
    }
#ifdef MDEBUG
    jlog("STAT: wchmm (re)build completed\n");
#endif
  }
  return(TRUE);
}

static boolean
multigram_append_to_global(DFA_INFO *gdfa, WORD_INFO *gwinfo, MULTIGRAM *m)
{
  /* the new grammar 'm' will be appended to the last of gdfa and gwinfo */
  m->state_begin = gdfa->state_num;     /* initial state ID */
  m->cate_begin = gdfa->term_num;       /* initial terminal ID */
  m->word_begin = gwinfo->num;  /* initial word ID */
  
  /* append category ID and node number of src DFA */
  /* Julius allow multiple initial states: connect each initial node
     is not necesarry. */
  dfa_append(gdfa, m->dfa, m->state_begin, m->cate_begin);
  /* append words of src vocabulary to global winfo */
  if (voca_append(gwinfo, m->winfo, m->cate_begin, m->word_begin) == FALSE) {
    return FALSE;
  }
  /* append category->word mapping table */
  terminfo_append(&(gdfa->term), &(m->dfa->term), m->cate_begin, m->word_begin);
  /* append catergory-pair information */
  /* pause has already been considered on m->dfa, so just append here */
  if (cpair_append(gdfa, m->dfa, m->cate_begin) == FALSE) {
    return FALSE;
  }
  /* re-set noise entry by merging */
  if (dfa_pause_word_append(gdfa, m->dfa, m->cate_begin) == FALSE) {
    return FALSE;
  }

  jlog("STAT: Gram #%d %s: installed\n", m->id, m->name);

  return TRUE;
}

int
multigram_add(DFA_INFO *dfa, WORD_INFO *winfo, char *name, PROCESS_LM *lm)
{
  MULTIGRAM *new;

  /* allocate new gram */
  new = (MULTIGRAM *)mymalloc(sizeof(MULTIGRAM));
  if (name != NULL) {
    strncpy(new->name, name, MAXGRAMNAMELEN);
  } else {
    strncpy(new->name, "(no name)", MAXGRAMNAMELEN);
  }

  new->id = lm->gram_maxid;
  new->dfa = dfa;
  new->winfo = winfo;
  /* will be setup and activated after multigram_update() is called once */
  new->hook = MULTIGRAM_DEFAULT | MULTIGRAM_ACTIVATE;
  new->newbie = TRUE;           /* need to setup */
  new->active = FALSE;          /* default: inactive */

  /* the new grammar is now added to gramlist */
  new->next = lm->grammars;
  lm->grammars = new;

  jlog("STAT: Gram #%d %s registered\n", new->id, new->name);
  lm->gram_maxid++;

  return new->id;
}

boolean
multigram_delete(int delid, PROCESS_LM *lm)
{
  MULTIGRAM *m;
  for(m=lm->grammars;m;m=m->next) {
    if (m->id == delid) {
      m->hook |= MULTIGRAM_DELETE;
      jlog("STAT: Gram #%d %s: marked delete\n", m->id, m->name);
      break;
    }
  }
  if (! m) {
    jlog("STAT: Gram #%d: not found\n", delid);
    return FALSE;
  }
  return TRUE;
}

void
multigram_delete_all(PROCESS_LM *lm)
{
  MULTIGRAM *m;
  for(m=lm->grammars;m;m=m->next) {
    m->hook |= MULTIGRAM_DELETE;
  }
}

static boolean
multigram_exec_delete(PROCESS_LM *lm)
{
  MULTIGRAM *m, *mtmp, *mprev;
  boolean ret_flag = FALSE;

  /* exec delete */
  mprev = NULL;
  m = lm->grammars;
  while(m) {
    mtmp = m->next;
    if (m->hook & MULTIGRAM_DELETE) {
      /* if any grammar is deleted, we need to rebuild lexicons etc. */
      /* so tell it to the caller */
      if (! m->newbie) ret_flag = TRUE;
      if (m->dfa) dfa_info_free(m->dfa);
      word_info_free(m->winfo);
      jlog("STAT: Gram #%d %s: purged\n", m->id, m->name);
      free(m);
      if (mprev != NULL) {
        mprev->next = mtmp;
      } else {
        lm->grammars = mtmp;
      }
    } else {
      mprev = m;
    }
    m = mtmp;
  }

  return(ret_flag);
}

int
multigram_activate(int gid, PROCESS_LM *lm)     /* only mark */
{
  MULTIGRAM *m;
  int ret;

  for(m=lm->grammars;m;m=m->next) {
    if (m->id == gid) {
      if (m->hook & MULTIGRAM_DEACTIVATE) {
        ret = 0;
        m->hook &= ~(MULTIGRAM_DEACTIVATE);
        m->hook |= MULTIGRAM_ACTIVATE;
        jlog("STAT: Gram #%d %s: marked active, superceding deactivate\n", m->id, m->name);
      } else {
        if (m->hook & MULTIGRAM_ACTIVATE) {
          jlog("STAT: Gram #%d %s: already marked active\n", m->id, m->name);
          ret = 1;
        } else {
          ret = 0;
          m->hook |= MULTIGRAM_ACTIVATE;
          jlog("STAT: Gram #%d %s: marked activate\n", m->id, m->name);
        }
      }
      break;
    }
  }
  if (! m) {
    jlog("WARNING: Gram #%d: not found, activation ignored\n", gid);
    ret = -1;
  }

  return(ret);
}

int
multigram_deactivate(int gid, PROCESS_LM *lm)   /* only mark */
{
  MULTIGRAM *m;
  int ret;

  for(m=lm->grammars;m;m=m->next) {
    if (m->id == gid) {
      if (m->hook & MULTIGRAM_ACTIVATE) {
        ret = 0;
        m->hook &= ~(MULTIGRAM_ACTIVATE);
        m->hook |= MULTIGRAM_DEACTIVATE;
        jlog("STAT: Gram #%d %s: marked deactivate, superceding activate\n", m->id, m->name);
      } else {
        if (m->hook & MULTIGRAM_DEACTIVATE) {
          jlog("STAT: Gram #%d %s: already marked deactivate\n", m->id, m->name);
          ret = 1;
        } else {
          ret = 0;
          m->hook |= MULTIGRAM_DEACTIVATE;
          jlog("STAT: Gram #%d %s: marked deactivate\n", m->id, m->name);
        }
      }
      break;
    }
  }
  if (! m) {
    jlog("WARNING: - Gram #%d: not found, deactivation ignored\n", gid);
    ret = -1;
  }

  return(ret);
}

static boolean
multigram_exec_activate(PROCESS_LM *lm)
{
  MULTIGRAM *m;
  boolean modified;
  
  modified = FALSE;
  for(m=lm->grammars;m;m=m->next) {
    if (m->hook & MULTIGRAM_ACTIVATE) {
      m->hook &= ~(MULTIGRAM_ACTIVATE);
      if (!m->active) {
        jlog("STAT: Gram #%d %s: turn on active\n", m->id, m->name);
      }
      m->active = TRUE;
      modified = TRUE;
    } else if (m->hook & MULTIGRAM_DEACTIVATE) {
      m->hook &= ~(MULTIGRAM_DEACTIVATE);
      if (m->active) {
        jlog("STAT: Gram #%d %s: turn off inactive\n", m->id, m->name);
      }
      m->active = FALSE;
      modified = TRUE;
    }
  }
  return(modified);
}
 
boolean                         /* return FALSE if no gram */
multigram_update(PROCESS_LM *lm)
{
  MULTIGRAM *m;
  boolean active_changed = FALSE;
  boolean rebuild_flag;

  if (lm->lmvar == LM_DFA_GRAMMAR) {
    /* setup additional grammar info of new ones */
    for(m=lm->grammars;m;m=m->next) {
      if (m->newbie) {
        jlog("STAT: Gram #%d %s: new grammar loaded, now mash it up for recognition\n", m->id, m->name);
        /* map dict item to dfa terminal symbols */
        if (make_dfa_voca_ref(m->dfa, m->winfo) == FALSE) {
          jlog("ERROR: failed to map dict <-> DFA. This grammar will be deleted\n");
          /* mark as to be deleted */
          m->hook |= MULTIGRAM_DELETE;
          continue;
        } 
        /* set dfa->sp_id and dfa->is_sp */
        dfa_find_pause_word(m->dfa, m->winfo, lm->am->hmminfo);
        /* build catergory-pair information */
        jlog("STAT: Gram #%d %s: extracting category-pair constraint for the 1st pass\n", m->id, m->name);
        if (extract_cpair(m->dfa) == FALSE) {
          jlog("ERROR: failed to extract category pair. This grammar will be deleted\n");
          /* mark as to be deleted */
          m->hook |= MULTIGRAM_DELETE;
        }
      }
    }
  }

  rebuild_flag = FALSE;
  /* delete grammars marked as "delete" */
  if (multigram_exec_delete(lm)) { /* some built grammars deleted */
    rebuild_flag = TRUE;        /* needs rebuilding global grammar */
  }
  /* find modified grammar */
  for(m=lm->grammars;m;m=m->next) {
    if (m->hook & MULTIGRAM_MODIFIED) {
      rebuild_flag = TRUE;      /* needs rebuilding global grammar */
      m->hook &= ~(MULTIGRAM_MODIFIED);
    }
  }

  if (rebuild_flag) {
    /* rebuild global grammar from scratch (including new) */
    /* active status not changed here (inactive grammar will also included) */
    /* activate/deactivate hook will be handled later, so just keep it here */
#ifdef MDEBUG
    jlog("STAT: re-build whole global grammar...\n");
#endif
    /* free old if not yet */
    if (lm->dfa != NULL) {
      dfa_info_free(lm->dfa);
      lm->dfa = NULL;
    }
    if (lm->winfo != NULL) {
      word_info_free(lm->winfo);
      lm->winfo = NULL;
    }
    /* concatinate all existing grammars to global */
    for(m=lm->grammars;m;m=m->next) {
      if (lm->lmvar == LM_DFA_GRAMMAR && lm->dfa == NULL) {
        lm->dfa = dfa_info_new();
        dfa_state_init(lm->dfa);
      }
      if (lm->winfo == NULL) {
        lm->winfo = word_info_new();
        winfo_init(lm->winfo);
      }
      if (m->newbie) m->newbie = FALSE;
      if (lm->lmvar == LM_DFA_WORD) {
        /* just append dictionaty */
        m->word_begin = lm->winfo->num;
        if (voca_append(lm->winfo, m->winfo, m->id, m->word_begin) == FALSE) {
          jlog("ERROR: multi-gram: failed to add dictionary #%d to recognition network\n", m->id);
          /* mark as delete */
          m->hook |= MULTIGRAM_DELETE;
        }
      } else {
        if (multigram_append_to_global(lm->dfa, lm->winfo, m) == FALSE) {
          jlog("ERROR: multi-gram: failed to add grammar #%d to recognition network\n", m->id);
          /* mark as delete */
          m->hook |= MULTIGRAM_DELETE;
        }
      }
    }
    /* delete the error grammars if exist */
    if (multigram_exec_delete(lm)) {
      jlog("ERROR: errorous grammar deleted\n");
    }
    lm->global_modified = TRUE;
  } else {                      /* global not need changed by the deletion */
    /* append only new grammars */
    for(m=lm->grammars;m;m=m->next) {
      if (m->newbie) {
        if (lm->lmvar == LM_DFA_GRAMMAR && lm->dfa == NULL) {
          lm->dfa = dfa_info_new();
          dfa_state_init(lm->dfa);
        }
        if (lm->winfo == NULL) {
          lm->winfo = word_info_new();
          winfo_init(lm->winfo);
        }
        if (m->newbie) m->newbie = FALSE;
        if (lm->lmvar == LM_DFA_WORD) {
          /* just append dictionaty */
          m->word_begin = lm->winfo->num;
          if (voca_append(lm->winfo, m->winfo, m->id, m->word_begin) == FALSE) {
            jlog("ERROR: multi-gram: failed to add dictionary #%d to recognition network\n", m->id);
            /* mark as delete */
            m->hook |= MULTIGRAM_DELETE;
          }
        } else {
          if (multigram_append_to_global(lm->dfa, lm->winfo, m) == FALSE) {
            jlog("ERROR: multi-gram: failed to add grammar #%d to recognition network\n", m->id);
            /* mark as delete */
            m->hook |= MULTIGRAM_DELETE;
          }
        }
        lm->global_modified = TRUE;
      }
    }
  }

  /* process activate/deactivate hook */
  active_changed = multigram_exec_activate(lm);

  if (lm->global_modified) {            /* if global lexicon has changed */
    /* now global grammar info has been updated */
    /* check if no grammar */
    if (lm->lmvar == LM_DFA_GRAMMAR) {
      if (lm->dfa == NULL || lm->winfo == NULL) {
        if (lm->dfa != NULL) {
          dfa_info_free(lm->dfa);
          lm->dfa = NULL;
        }
        if (lm->winfo != NULL) {
          word_info_free(lm->winfo);
          lm->winfo = NULL;
        }
      }
    }
#ifdef MDEBUG
    jlog("STAT: grammar update completed\n");
#endif
  }

  if (lm->global_modified || active_changed) {
    return (TRUE);
  }

  return FALSE;
}

static boolean
multigram_read_file_and_add(char *dfa_file, char *dict_file, PROCESS_LM *lm)
{
  WORD_INFO *new_winfo;
  DFA_INFO *new_dfa;
  char buf[MAXGRAMNAMELEN], *p, *q;
  boolean ret;

  if (dfa_file != NULL) {
    jlog("STAT: reading [%s] and [%s]...\n", dfa_file, dict_file);
  } else {
    jlog("STAT: reading [%s]...\n", dict_file);
  }
  
  /* read dict*/
  new_winfo = word_info_new();

  if (lm->lmvar == LM_DFA_GRAMMAR) {
    ret = init_voca(new_winfo, dict_file, lm->am->hmminfo, 
#ifdef MONOTREE
                    TRUE,
#else 
                    FALSE,
#endif
                    lm->config->forcedict_flag);
    if ( ! ret ) {
      jlog("ERROR: failed to read dictionary \"%s\"\n", dict_file);
      word_info_free(new_winfo);
      return FALSE;
    }
  } else if (lm->lmvar == LM_DFA_WORD) {
    ret = init_wordlist(new_winfo, dict_file, lm->am->hmminfo, 
                        lm->config->wordrecog_head_silence_model_name,
                        lm->config->wordrecog_tail_silence_model_name,
                        (lm->config->wordrecog_silence_context_name[0] == '\0') ? NULL : lm->config->wordrecog_silence_context_name,
                        lm->config->forcedict_flag);
    if ( ! ret ) {
      jlog("ERROR: failed to read word list \"%s\"\n", dict_file);
      word_info_free(new_winfo);
      return FALSE;
    }
  }

  new_dfa = NULL;
  if (lm->lmvar == LM_DFA_GRAMMAR) {
    /* read dfa */
    new_dfa = dfa_info_new();
    if (init_dfa(new_dfa, dfa_file) == FALSE) {
      jlog("ERROR: multi-gram: error in reading DFA\n");
      word_info_free(new_winfo);
      dfa_info_free(new_dfa);
      return FALSE;
    }
  }

  jlog("STAT: done\n");

  /* extract name */
  p = &(dict_file[0]);
  q = p;
  while(*p != '\0') {
    if (*p == '/') q = p + 1;
    p++;
  }
  p = q;
  while(*p != '\0' && *p != '.') {
    buf[p-q] = *p;
    p++;
  }
  buf[p-q] = '\0';
  
  /* register the new grammar to multi-gram tree */
  multigram_add(new_dfa, new_winfo, buf, lm);

  return TRUE;

}


boolean
multigram_load_all_gramlist(PROCESS_LM *lm)
{
  GRAMLIST *g;
  GRAMLIST *groot;
  boolean ok_p;

  switch(lm->config->lmvar) {
  case LM_DFA_GRAMMAR: groot = lm->config->gramlist_root; break;
  case LM_DFA_WORD:    groot = lm->config->wordlist_root; break;
  }

  ok_p = TRUE;
  for(g = groot; g; g = g->next) {
    if (multigram_read_file_and_add(g->dfafile, g->dictfile, lm) == FALSE) {
      ok_p = FALSE;
    }
  }
  return(ok_p);
}

int
multigram_get_all_num(PROCESS_LM *lm)
{
  MULTIGRAM *m;
  int cnt;
  
  cnt = 0;
  for(m=lm->grammars;m;m=m->next) cnt++;
  return(cnt);
}

int
multigram_get_gram_from_category(int category, PROCESS_LM *lm)
{
  MULTIGRAM *m;
  int tb, te;
  for(m = lm->grammars; m; m = m->next) {
    if (m->newbie) continue;
    tb = m->cate_begin;
    te = tb + m->dfa->term_num;
    if (tb <= category && category < te) { /* found */
      return(m->id);
    }
  }
  return(-1);
}

int
multigram_get_gram_from_wid(WORD_ID wid, PROCESS_LM *lm)
{
  MULTIGRAM *m;
  int wb, we;

  for(m = lm->grammars; m; m = m->next) {
    if (m->newbie) continue;
    wb = m->word_begin;
    we = wb + m->winfo->num;
    if (wb <= wid && wid < we) { /* found */
      return(m->id);
    }
  }
  return(-1);
}


void
multigram_free_all(MULTIGRAM *root)
{
  MULTIGRAM *m, *mtmp;

  m = root;
  while(m) {
    mtmp = m->next;
    if (m->dfa) dfa_info_free(m->dfa);
    word_info_free(m->winfo);
    free(m);
    m = mtmp;
  }
}

int
multigram_get_id_by_name(PROCESS_LM *lm, char *gramname)
{
  MULTIGRAM *m;

  for(m=lm->grammars;m;m=m->next) {
    if (strmatch(m->name, gramname)) break;
  }
  if (!m) {
    jlog("ERROR: multigram: cannot find grammar \"%s\"\n", gramname);
    return -1;
  }
  return m->id;
}

MULTIGRAM *
multigram_get_grammar_by_name(PROCESS_LM *lm, char *gramname)
{
  MULTIGRAM *m;

  for(m=lm->grammars;m;m=m->next) {
    if (strmatch(m->name, gramname)) break;
  }
  if (!m) {
    jlog("ERROR: multigram: cannot find grammar \"%s\"\n", gramname);
    return NULL;
  }
  return m;
}

MULTIGRAM *
multigram_get_grammar_by_id(PROCESS_LM *lm, unsigned short id)
{
  MULTIGRAM *m;

  for(m=lm->grammars;m;m=m->next) {
    if (m->id == id) break;
  }
  if (!m) {
    jlog("ERROR: multi-gram: cannot find grammar id \"%d\"\n", id);
    return NULL;
  }
  return m;
}

boolean
multigram_add_words_to_grammar(PROCESS_LM *lm, MULTIGRAM *m, WORD_INFO *winfo)
{
  int offset;

  if (lm == NULL || m == NULL || winfo == NULL) return FALSE;

  offset = m->winfo->num;
  printf("adding %d words to grammar #%d (%d words)\n", winfo->num, m->id, m->winfo->num);
  /* append to the grammar */
  if (voca_append(m->winfo, winfo, m->id, offset) == FALSE) {
    jlog("ERROR: multi-gram: failed to add words to dict in grammar #%d \"%s\"\n", m->id, m->name);
    return FALSE;
  }
  /* update dictianary info */
  if (lm->lmvar == LM_DFA_GRAMMAR) {
    if (m->dfa->term_num != 0) free_terminfo(&(m->dfa->term));
    if (make_dfa_voca_ref(m->dfa, m->winfo) == FALSE) {
      jlog("ERROR: failed to map dict <-> DFA. This grammar will be deleted\n");
      return FALSE;
    } 
  }
  /* prepare for update */
  m->hook |= MULTIGRAM_MODIFIED;

  return TRUE;
}

boolean
multigram_add_words_to_grammar_by_name(PROCESS_LM *lm, char *gramname, WORD_INFO *winfo)
{
  return(multigram_add_words_to_grammar(lm, multigram_get_grammar_by_name(lm, gramname), winfo));
}

boolean
multigram_add_words_to_grammar_by_id(PROCESS_LM *lm, unsigned short id, WORD_INFO *winfo)
{
  return(multigram_add_words_to_grammar(lm, multigram_get_grammar_by_id(lm, id), winfo));
}

/* end of file */