void load_context_cue(arpa_lm_t* lm, char* ccs_filename)
{
FILE* context_cues_fp;
char wlist_entry[1024];
char current_cc[200];
vocab_sz_t current_cc_id;
lm->context_cue = (flag *) rr_calloc(lm->table_sizes[0],sizeof(flag));
lm->no_of_ccs = 0;
if (strcmp(ccs_filename,"")) {
context_cues_fp = rr_iopen(ccs_filename);
while (fgets (wlist_entry, sizeof (wlist_entry),context_cues_fp)) {
if (strncmp(wlist_entry,"##",2)==0) continue;
sscanf (wlist_entry, "%s ",current_cc);
warn_on_wrong_vocab_comments(wlist_entry);
if (sih_lookup(lm->vocab_ht,current_cc,¤t_cc_id) == 0)
quit(-1,"Error : %s in the context cues file does not appear in the vocabulary.\n",current_cc);
lm->context_cue[(unsigned short) current_cc_id] = 1;
lm->no_of_ccs++;
fprintf(stderr,"Context cue word : %s id = %lld\n",current_cc,current_cc_id);
}
rr_iclose(context_cues_fp);
}
}
void calc_mem_req(ng_t *ng,flag is_ascii) {
ngram current_ngram;
ngram previous_ngram;
count_t *ng_count;
int i,j;
current_ngram.id_array = (id__t *) rr_malloc(sizeof(id__t)*ng->n);
previous_ngram.id_array = (id__t *) rr_malloc(sizeof(id__t)*ng->n);
ng_count = (count_t *) rr_calloc(ng->n,sizeof(count_t));
current_ngram.n = ng->n;
rewind(ng->id_gram_fp);
while (!rr_feof(ng->id_gram_fp)) {
for (i=0;i<=ng->n-1;i++) {
previous_ngram.id_array[i]=current_ngram.id_array[i];
}
get_ngram(ng->id_gram_fp,¤t_ngram,is_ascii);
for (i=0;i<=ng->n-1;i++) {
if (current_ngram.id_array[i] != previous_ngram.id_array[i]) {
for (j=i;j<=ng->n-1;j++) {
if (j>0) {
if (ng_count[j] > ng->cutoffs[j-1]) {
ng->table_sizes[j]++;
}
}
ng_count[j] = current_ngram.count;
}
i=ng->n;
}
else {
ng_count[i] += current_ngram.count;
}
}
}
for (i=1;i<=ng->n-1;i++) {
if (ng_count[i] > ng->cutoffs[i-1]) {
ng->table_sizes[i]++;
}
}
/* Add a fudge factor, as problems can crop up with having to
cut-off last few n-grams. */
for (i=1;i<=ng->n-1;i++) {
ng->table_sizes[i]+=10;
}
rr_iclose(ng->id_gram_fp);
ng->id_gram_fp = rr_iopen(ng->id_gram_filename);
}
void read_vocab(ng_t* ng, int verbosity)
{
vocab_sz_t test_cc_id;
vocab_sz_t current_cc_id;
char current_cc[200];
char wlist_entry[1024];
pc_message(verbosity,2,"Reading vocabulary.\n");
/* Don't change the parameter of sih_create, because it will change
the binary layout of the .binlm file */
ng->vocab_ht =
sih_create(1000,0.5,2.0,1);
read_voc(ng->vocab_filename,verbosity,ng->vocab_ht,&ng->vocab,&(ng->vocab_size));
/* Determine which of the vocabulary words are context cues */
ng->no_of_ccs = 0;
ng->context_cue = (flag *) rr_calloc(ng->vocab_size+1,sizeof(flag));
if (ng->context_set) {
/* This should be tied to l889 to l894 in lm_combine.c
*/
while (fgets (wlist_entry, sizeof (wlist_entry),ng->context_cues_fp)) {
if (strncmp(wlist_entry,"##",2)==0) continue;
sscanf (wlist_entry, "%s ",current_cc);
warn_on_wrong_vocab_comments(wlist_entry);
if (sih_lookup(ng->vocab_ht,current_cc,¤t_cc_id) == 0)
pc_message(verbosity,1,"Warning : %s in the context cues file does not appear in the vocabulary.\n",current_cc);
else {
ng->context_cue[(unsigned short) current_cc_id] = 1;
pc_message(verbosity,2,"Context cue word : %s id = %d\n",current_cc,current_cc_id);
ng->no_of_ccs++;
}
}
rr_iclose(ng->context_cues_fp);
}
if ((sih_lookup(ng->vocab_ht,"<s>",&test_cc_id) != 0))
if (ng->context_cue[(unsigned short) test_cc_id] == 0)
fprintf(stderr,"WARNING: <s> appears as a vocabulary item, but is not labelled as a\ncontext cue.\n");
if ((sih_lookup(ng->vocab_ht,"<p>",&test_cc_id) != 0))
if (ng->context_cue[(unsigned short) test_cc_id] == 0)
fprintf(stderr,"WARNING: <p> appears as a vocabulary item, but is not labelled as a\ncontext cue.\n");
if ((sih_lookup(ng->vocab_ht,"<art>",&test_cc_id) != 0))
if (ng->context_cue[(unsigned short) test_cc_id] == 0)
fprintf(stderr,"WARNING: <art> appears as a vocabulary item, but is not labelled as a\ncontext cue.\n");
}
void load_weights(double* w1, double* w2,char* file)
{
FILE* fp;
fp=rr_iopen(file);
if (fscanf(fp,"%*s%lf",w1)!=1)
quit(-1,"Error in reading weight file\n");
if (fscanf(fp,"%*s%lf",w2)!=1)
quit(-1,"Error in reading weight file\n");
rr_iclose(fp);
}
void read_wlist_into_siht(char *wlist_filename, int verbosity,
sih_t *p_word_id_ht, /** Pointer of the word ID hash table */
vocab_sz_t * p_n_wlist /** Pointer of the size of wordlist */
)
{
static char rname[]="read_wlist_into_siht";
FILE *wlist_fp = rr_iopen(wlist_filename);
char wlist_entry[1024], word[256], *word_copy;
vocab_sz_t entry_no = 0;
while (fgets (wlist_entry, sizeof (wlist_entry), wlist_fp)) {
if (strncmp(wlist_entry,"##",2) == 0) continue;
entry_no++;
/* printf("entry no %lld, wlist_entry %s\n",entry_no,wlist_entry);*/
if(entry_no%1000==0){
fprintf(stdout,".");
fflush(stdout);
}
sscanf (wlist_entry, "%s ", word);
warn_on_wrong_vocab_comments(wlist_entry);
word_copy = rr_salloc(word);
sih_add(p_word_id_ht, word_copy, entry_no);
free(word_copy); // HLW
}
fprintf(stdout,"\n");
fflush(stdout);
rr_iclose(wlist_fp);
if(verbose_cmuclmtk == 1) {
if (verbosity)
fprintf(stderr,"%s: a list of %d words was read from \"%s\".\n",
rname,(int) entry_no,wlist_filename);
}
*p_n_wlist = entry_no;
}
void read_voc(char *filename, int verbosity,
sih_t *p_vocab_ht, char ***p_vocab,
unsigned short *p_vocab_size)
/* p_vocab==NULL means: build only a hash table */
{
/* static char rname[] = "rd_voc"; */ /* Never used anyway! */
char *pperiod;
int vocab_size;
pperiod = rindex(filename,'.');
if (pperiod==NULL) pperiod = filename-1;
if (strcmp(pperiod+1,"vocab_ht")==0) { /* file == hash_table */
FILE *fp=rr_iopen(filename);
sih_val_read_from_file(p_vocab_ht, fp, filename, verbosity);
rr_iclose(fp);
vocab_size = p_vocab_ht->nentries;
if (p_vocab!=NULL) {
get_vocab_from_vocab_ht(p_vocab_ht, vocab_size, verbosity, p_vocab);
*p_vocab[0] = salloc("<UNK>");
}
}
else { /* file == vocab(ascii) */
read_wlist_into_siht(filename, verbosity, p_vocab_ht, &vocab_size);
if (p_vocab!=NULL) {
read_wlist_into_array(filename, verbosity, p_vocab, &vocab_size);
*p_vocab[0] = salloc("<UNK>");
}
}
if (p_vocab_size) {
*p_vocab_size = vocab_size;
}
}
void compute_perplexity(ng_t *ng,
arpa_lm_t *arpa_ng,
char *text_stream_filename,
char *probs_stream_filename,
char *annotation_filename,
char *oov_filename,
char *fb_list_filename,
flag backoff_from_unk_inc,
flag backoff_from_unk_exc,
flag backoff_from_ccs_inc,
flag backoff_from_ccs_exc,
flag arpa_lm,
flag include_unks,
double log_base) {
fb_info *fb_list;
FILE *temp_fp;
FILE *text_stream_fp;
FILE *probs_stream_fp;
FILE *annotation_fp;
FILE *oov_fp;
flag out_probs;
flag annotate;
flag out_oovs;
flag found_unk_wrongly;
double prob;
double sum_log_prob;
int total_words;
int excluded_unks;
int excluded_ccs;
char current_word[1000]; /* Hope that's big enough */
char **prev_words;
vocab_sz_t current_id;
id__t short_current_id;
id__t *context;
int context_length;
int i;
int bo_case;
int actual_context_length;
int *ngrams_hit;
int n;
/* Initialise file pointers to prevent warnings from the compiler. */
probs_stream_fp = NULL;
annotation_fp = NULL;
oov_fp = NULL;
short_current_id = 0;
found_unk_wrongly = 0;
annotate = 0;
bo_case = 0;
if (arpa_lm) {
n = arpa_ng->n;
fb_list = gen_fb_list(arpa_ng->vocab_ht,
(int) arpa_ng->vocab_size,
arpa_ng->vocab,
arpa_ng->context_cue,
backoff_from_unk_inc,
backoff_from_unk_exc,
backoff_from_ccs_inc,
backoff_from_ccs_exc,
fb_list_filename);
}else {
n = ng->n;
fb_list = gen_fb_list(ng->vocab_ht,
(int) ng->vocab_size,
ng->vocab,
ng->context_cue,
backoff_from_unk_inc,
backoff_from_unk_exc,
backoff_from_ccs_inc,
backoff_from_ccs_exc,
fb_list_filename);
}
ngrams_hit = (int *) rr_calloc(n,sizeof(int));
prev_words = (char **) rr_malloc(sizeof(char *)*n);
for (i=0;i<=n-1;i++)
prev_words[i] = (char *) rr_malloc(sizeof(char)*1000);
/* Check that text_stream_filename and probs_stream_filename (if
specified) are valid. Note that the checks employed by the
standard rr_fopen tools are not suitable here, since we don't
want the program to terminate if the paths are not found. */
if (!strcmp(text_stream_filename,"")) {
printf("Error : Must specify a text file. Use the -text switch.\n");
return;
}
if (!rr_fexists(text_stream_filename) && strcmp(text_stream_filename,"-")) {
printf("Error : Can't open file %s for reading.\n",text_stream_filename);
return;
}
out_probs = strcmp(probs_stream_filename,"");
annotate = strcmp(annotation_filename,"");
out_oovs = strcmp(oov_filename,"");
printf("Computing perplexity of the language model with respect\n");
printf(" to the text %s\n",text_stream_filename);
if (out_probs)
printf("Probability stream will be written to file %s\n",
probs_stream_filename);
if (annotate)
printf("Annotation will be written to file %s\n",
annotation_filename);
if (out_oovs)
printf("Out of vocabulary words will be written to file %s\n",
oov_filename);
if (backoff_from_unk_inc)
printf("Will force inclusive back-off from OOVs.\n");
if (backoff_from_unk_exc)
printf("Will force exclusive back-off from OOVs.\n");
if (backoff_from_ccs_inc)
printf("Will force inclusive back-off from context cues.\n");
if (backoff_from_ccs_exc)
printf("Will force exclusive back-off from context cues.\n");
if (strcmp(fb_list_filename,""))
printf("Will force back-off according to the contents of %s\n",
fb_list_filename);
if (include_unks)
printf("Perplexity calculation will include OOVs.\n");
/* Check for existance of files, as rr functions will quit, which isn't
what we want */
if (out_probs && strcmp(probs_stream_filename,"-")) {
if ((temp_fp = fopen(probs_stream_filename,"w")) == NULL) {
printf("Error : Can't open file %s for writing.\n",probs_stream_filename);
return;
}
fclose(temp_fp);
}
if (annotate && strcmp(annotation_filename,"-")) {
if ((temp_fp = fopen(annotation_filename,"w")) == NULL) {
printf("Error : Can't open file %s for writing.\n",annotation_filename);
return;
}
fclose(temp_fp);
}
if (out_oovs && strcmp(oov_filename,"-")) {
if ((temp_fp = fopen(oov_filename,"w")) == NULL) {
printf("Error : Can't open file %s for writing.\n",oov_filename);
return;
}
fclose(temp_fp);
}
text_stream_fp = rr_iopen(text_stream_filename);
if (out_probs)
probs_stream_fp = rr_oopen(probs_stream_filename);
if (annotate)
annotation_fp = rr_oopen(annotation_filename);
if (out_oovs)
oov_fp = rr_oopen(oov_filename);
context = (id__t *) rr_malloc(sizeof(id__t)*(n-1));
sum_log_prob = 0.0;
total_words = 0;
excluded_unks = 0;
excluded_ccs = 0;
while (!rr_feof(text_stream_fp)) {
if (total_words > 0) {
if (total_words < n)
strcpy(prev_words[total_words-1],current_word);
else {
for (i=0;i<=n-3;i++)
strcpy(prev_words[i],prev_words[i+1]);
if (n>1)
strcpy(prev_words[n-2],current_word);
}
}
if (total_words < (n-1))
context_length = total_words;
else
context_length = n-1;
/* Fill context with right stuff */
if (total_words > (n-1)) {
for (i=0;i<=context_length-2;i++)
context[i] = context[i+1];
}
if (context_length != 0)
context[context_length-1] = short_current_id;
if (fscanf(text_stream_fp,"%s",current_word) != 1) {
if (!rr_feof(text_stream_fp)) {
printf("Error reading text file.\n");
return;
}
}
if (!rr_feof(text_stream_fp)) {
if (arpa_lm) {
sih_lookup(arpa_ng->vocab_ht,current_word,¤t_id);
if (arpa_ng->vocab_type == CLOSED_VOCAB && current_id == 0) {
found_unk_wrongly = 1;
printf("Error : %s is not in the vocabulary, and this is a closed \nvocabulary model.\n",current_word);
}
if (current_id > arpa_ng->vocab_size)
quit(-1,"Error : returned value from sih_lookup (%d) is too high.\n",context[i]);
else
short_current_id = current_id;
}else {
sih_lookup(ng->vocab_ht,current_word,¤t_id);
if (ng->vocab_type == CLOSED_VOCAB && current_id == 0) {
found_unk_wrongly = 1;
printf("Error : %s is not in the vocabulary, and this is a closed \nvocabulary model.\n",current_word);
}
if (current_id > ng->vocab_size)
quit(-1,"Error : returned value from sih_lookup (%d) is too high.\n",context[i]);
else
short_current_id = current_id;
}
if (!found_unk_wrongly) {
if (current_id == 0 && out_oovs)
fprintf(oov_fp,"%s\n",current_word);
if ((arpa_lm && (!(arpa_ng->context_cue[current_id])))
|| ((!arpa_lm) && (!(ng->context_cue[current_id])))) {
if (include_unks || current_id != 0) {
prob = calc_prob_of(short_current_id,
context,
context_length,
ng,
arpa_ng,
fb_list,
&bo_case,
&actual_context_length,
arpa_lm);
if (prob<= 0.0 || prob > 1.0) {
fprintf(stderr,"Warning : ");
if (short_current_id == 0)
fprintf(stderr,"P( <UNK> | ");
else
fprintf(stderr,"P( %s | ",current_word);
for (i=0;i<=actual_context_length-1;i++) {
if (context[i+context_length-actual_context_length] == 0)
fprintf(stderr,"<UNK> ");
else
fprintf(stderr,"%s ",prev_words[i]);
}
fprintf(stderr,") = %g logprob = %g \n ",prob,log(prob)/log(log_base));
fprintf(stderr,"bo_case == 0x%dx, actual_context_length == %d\n",
bo_case, actual_context_length);
}
if (annotate) {
if (short_current_id == 0)
fprintf(annotation_fp,"P( <UNK> | ");
else
fprintf(annotation_fp,"P( %s | ",current_word);
for (i=0;i<=actual_context_length-1;i++) {
if (context[i+context_length-actual_context_length] == 0)
fprintf(annotation_fp,"<UNK> ");
else {
if (arpa_lm)
fprintf(annotation_fp,"%s ",arpa_ng->vocab[context[i+context_length-actual_context_length]]);
else
fprintf(annotation_fp,"%s ",ng->vocab[context[i+context_length-actual_context_length]]);
}
}
fprintf(annotation_fp,") = %g logprob = %f bo_case = ",prob,log(prob)/log(log_base));
decode_bo_case(bo_case,actual_context_length,annotation_fp);
}
/* Calculate level to which we backed off */
for (i=actual_context_length-1;i>=0;i--) {
int four_raise_i = 1<<(2*i); /* PWP */
/*
* PWP: This was "if ((bo_case / (int) pow(3,i)) == 0)"
* but was getting a divide-by-zero error on an Alpha
* (it isn't clear to me why it should ever have done so)
* Anyway, it is much faster to do in base-4.
*/
if ((bo_case == 0) || ((bo_case / four_raise_i) == 0)) {
ngrams_hit[i+1]++;
i = -2;
}else
bo_case -= ((bo_case / four_raise_i) * four_raise_i);
}
if (i != -3)
ngrams_hit[0]++;
if (out_probs)
fprintf(probs_stream_fp,"%g\n",prob);
sum_log_prob += log10(prob);
}
if (current_id == 0 && !include_unks)
excluded_unks++;
}
else {
if (((!arpa_lm) && ng->context_cue[current_id]) ||
(arpa_lm && arpa_ng->context_cue[current_id]))
excluded_ccs++;
}
total_words++;
}
}
}
if (!found_unk_wrongly) { /* pow(x,y) = e**(y ln(x)) */
printf("Perplexity = %.2f, Entropy = %.2f bits\n",
exp(-sum_log_prob/(total_words-excluded_ccs-excluded_unks) *
log(10.0)),
(-sum_log_prob/(total_words-excluded_ccs-excluded_unks) *
log(10.0) / log(2.0)));
printf("Computation based on %d words.\n",
total_words-excluded_ccs-excluded_unks);
for(i=n;i>=1;i--) {
printf("Number of %d-grams hit = %d (%.2f%%)\n",i,ngrams_hit[i-1],
(float) 100*ngrams_hit[i-1]/(total_words-excluded_ccs-excluded_unks) );
}
printf("%d OOVs (%.2f%%) and %d context cues were removed from the calculation.\n",
excluded_unks,
(float) 100*excluded_unks/(total_words-excluded_ccs),excluded_ccs);
}
rr_iclose(text_stream_fp);
if (out_probs)
rr_oclose(probs_stream_fp);
if (annotate)
rr_oclose(annotation_fp);
if (out_oovs)
rr_oclose(oov_fp);
free (fb_list);
free (context);
free (ngrams_hit);
}
int oe_02_main (int argc, char **argv) {
ng_t ng;
arpa_lm_t arpa_ng;
char input_string[500];
int num_of_args;
char *args[MAX_ARGS];
char *lm_filename_arpa;
char *lm_filename_binary;
flag told_to_quit;
flag inconsistant_parameters;
flag backoff_from_unk_inc;
flag backoff_from_unk_exc;
flag backoff_from_ccs_inc;
flag backoff_from_ccs_exc;
flag arpa_lm;
flag binary_lm;
flag include_unks;
char *fb_list_filename;
char *probs_stream_filename;
char *annotation_filename;
char *text_stream_filename;
char *oov_filename;
char *ccs_filename;
int generate_size;
int random_seed;
double log_base;
char wlist_entry[1024];
char current_cc[200];
vocab_sz_t current_cc_id;
FILE *context_cues_fp;
int n;
/* Process command line */
report_version(&argc,argv);
if (pc_flagarg(&argc, argv,"-help") ||
argc == 1 ||
(strcmp(argv[1],"-binary") && strcmp(argv[1],"-arpa"))) {
oe_02_help_message();
exit(1);
}
lm_filename_arpa = rr_salloc(pc_stringarg(&argc, argv,"-arpa",""));
if (strcmp(lm_filename_arpa,""))
arpa_lm = 1;
else
arpa_lm = 0;
lm_filename_binary = rr_salloc(pc_stringarg(&argc, argv,"-binary",""));
if (strcmp(lm_filename_binary,""))
binary_lm = 1;
else
binary_lm = 0;
if (arpa_lm && binary_lm)
quit(-1,"Error : Can't use both -arpa and -binary flags.\n");
if (!arpa_lm && !binary_lm)
quit(-1,"Error : Must specify either a binary or an arpa format language model.\n");
ccs_filename = rr_salloc(pc_stringarg(&argc, argv,"-context",""));
if (binary_lm && strcmp(ccs_filename,""))
fprintf(stderr,"Warning - context cues file not needed with binary language model file.\nWill ignore it.\n");
pc_report_unk_args(&argc,argv,2);
/* Load language model */
if (arpa_lm) {
fprintf(stderr,"Reading in language model from file %s\n",
lm_filename_arpa);
load_arpa_lm(&arpa_ng,lm_filename_arpa);
}else {
fprintf(stderr,"Reading in language model from file %s\n",
lm_filename_binary);
load_lm(&ng,lm_filename_binary);
}
fprintf(stderr,"\nDone.\n");
n=arpa_lm?
arpa_ng.n:
ng.n;
if (arpa_lm) {
arpa_ng.context_cue =
(flag *) rr_calloc(arpa_ng.table_sizes[0],sizeof(flag));
arpa_ng.no_of_ccs = 0;
if (strcmp(ccs_filename,"")) {
context_cues_fp = rr_iopen(ccs_filename);
while (fgets (wlist_entry, sizeof (wlist_entry),context_cues_fp)) {
if (strncmp(wlist_entry,"##",2)==0) continue;
sscanf (wlist_entry, "%s ",current_cc);
warn_on_wrong_vocab_comments(wlist_entry);
if (sih_lookup(arpa_ng.vocab_ht,current_cc,¤t_cc_id) == 0)
quit(-1,"Error : %s in the context cues file does not appear in the vocabulary.\n",current_cc);
arpa_ng.context_cue[(unsigned short) current_cc_id] = 1;
arpa_ng.no_of_ccs++;
fprintf(stderr,"Context cue word : %s id = %lld\n",current_cc,current_cc_id);
}
rr_iclose(context_cues_fp);
}
}
/* Process commands */
told_to_quit = 0;
num_of_args = 0;
while (!feof(stdin) && !told_to_quit) {
printf("evallm : \n");
fgets(input_string, sizeof(input_string), stdin);
if(strlen(input_string) < sizeof(input_string)-1)
input_string[strlen(input_string)-1] = '\0'; //chop new-line
else
quit(1, "evallm input exceeds size of input buffer");
if (!feof(stdin)) {
parse_comline(input_string,&num_of_args,args);
log_base = pc_doublearg(&num_of_args,args,"-log_base",10.0);
backoff_from_unk_inc = pc_flagarg(&num_of_args,args,"-backoff_from_unk_inc");
backoff_from_ccs_inc = pc_flagarg(&num_of_args,args,"-backoff_from_ccs_inc");
backoff_from_unk_exc = pc_flagarg(&num_of_args,args,"-backoff_from_unk_exc");
backoff_from_ccs_exc = pc_flagarg(&num_of_args,args,"-backoff_from_ccs_exc");
include_unks = pc_flagarg(&num_of_args,args,"-include_unks");
fb_list_filename = rr_salloc(pc_stringarg(&num_of_args,args,"-backoff_from_list",""));
text_stream_filename =
rr_salloc(pc_stringarg(&num_of_args,args,"-text",""));
probs_stream_filename =
rr_salloc(pc_stringarg(&num_of_args,args,"-probs",""));
annotation_filename =
rr_salloc(pc_stringarg(&num_of_args,args,"-annotate",""));
oov_filename = rr_salloc(pc_stringarg(&num_of_args,args,"-oovs",""));
generate_size = pc_intarg(&num_of_args,args,"-size",10000);
random_seed = pc_intarg(&num_of_args,args,"-seed",-1);
inconsistant_parameters = 0;
if (backoff_from_unk_inc && backoff_from_unk_exc) {
fprintf(stderr,"Error : Cannot specify both exclusive and inclusive forced backoff.\n");
fprintf(stderr,"Use only one of -backoff_from_unk_exc and -backoff_from_unk_inc\n");
inconsistant_parameters = 1;
}
if (backoff_from_ccs_inc && backoff_from_ccs_exc) {
fprintf(stderr,"Error : Cannot specify both exclusive and inclusive forced backoff.\n");
fprintf(stderr,"Use only one of -backoff_from_ccs_exc and -backoff_from_ccs_inc\n");
inconsistant_parameters = 1;
}
if (num_of_args > 0) {
if (!inconsistant_parameters) {
if (!strcmp(args[0],"perplexity")) {
compute_perplexity(&ng,
&arpa_ng,
text_stream_filename,
probs_stream_filename,
annotation_filename,
oov_filename,
fb_list_filename,
backoff_from_unk_inc,
backoff_from_unk_exc,
backoff_from_ccs_inc,
backoff_from_ccs_exc,
arpa_lm,
include_unks,
log_base);
}else
/* do perplexity sentence by sentence [20090612] (air) */
if (!strcmp(args[0],"uttperp")) {
FILE *uttfh,*tempfh;
char utt[4096]; /* live dangerously... */
char tmpfil[128];
if ((uttfh = fopen(text_stream_filename,"r")) == NULL) {
printf("Error: can't open %s\n",text_stream_filename);
exit(1);
}
char *template = "uttperp_XXXXXX";// CHANGED HLW
mkstemp(template);// CHANGED HLW
int oe_15_main( int argc, char **argv )
{
FILE **fin;
ngram *ng;
ngram outng;
flag *done, finished;
int i, j, nfiles;
/* Process the command line */
report_version(&argc,argv);
procComLine( &argc, argv ) ;
if( argc < 2 ) {
printUsage( argv[0] ) ;
exit( 1 ) ;
}
nfiles = argc - 1;
/* allocate memory */
fin = (FILE **) rr_malloc( sizeof( FILE *) * nfiles );
done = (flag *) rr_malloc( sizeof( flag ) * nfiles );
ng = (ngram *) rr_malloc( sizeof( ngram ) * nfiles );
for( i = 0; i < nfiles; i++ ) {
ng[i].id_array = (id__t *) rr_calloc( n, sizeof( id__t ) );
ng[i].n = n;
}
outng.id_array = (id__t *) rr_calloc( n, sizeof( id__t ) );
outng.n = n;
/* open the input files */
for( i = 0; i < nfiles; i++ )
fin[i] = rr_iopen( argv[i+1] );
/* read first ngram from each file */
for( i = 0; i < nfiles; i++ ) {
done[i] = 0;
if ( !get_ngram( fin[i], &ng[i], ascii_in ) )
done[i] = 1;
}
finished = 0;
while ( !finished ) {
/* set outng to max possible */
for( i = 0; i < n; i++ )
outng.id_array[i] = MAX_VOCAB_SIZE;
/* find smallest ngram */
for( i = 0; i < nfiles; i++ ) {
if ( !done[i] )
if ( cmp_ngram( &outng, &ng[i] ) > 0 )
for( j = 0; j < n; j++ ) outng.id_array[j] = ng[i].id_array[j];
}
outng.count = 0;
for( i = 0; i < nfiles; i++ ) {
if ( !done[i] ) {
/* add counts of equal ngrams */
if ( cmp_ngram( &outng, &ng[i] ) == 0 ) {
outng.count += ng[i].count;
if ( !get_ngram( fin[i], &ng[i], ascii_in ) ) {
/* check if all files done */
done[i] = 1;
finished = 1;
for( j = 0; j < nfiles; j++ )
if ( ! done[j] ) finished = 0;
}
}
}
}
write_ngram( stdout, &outng, ascii_out );
}
for( i = 0; i < nfiles; i++ )
rr_iclose( fin[i] );
fprintf(stderr,"mergeidngram : Done.\n");
return( 0 );
}
void merge_tempfiles (int start_file,
int end_file,
char *temp_file_root,
char *temp_file_ext,
int max_files,
FILE *outfile,
int n,
int verbosity) {
FILE *new_temp_file;
char *new_temp_filename;
FILE **temp_file;
char **temp_filename;
char **current_ngram;
char smallest_ngram[1000];
int *current_ngram_count;
flag *finished;
flag all_finished;
int temp_count;
char temp_word[500];
int i,j;
pc_message(verbosity,2,"Merging temp files %d through %d...\n", start_file,
end_file);
/*
* If we try to do more than max_files, then merge into groups,
* then merge groups recursively.
*/
if (end_file-start_file+1 > max_files) {
int new_start_file, new_end_file;
int n_file_groups = 1 + (end_file-start_file)/max_files;
fprintf(stderr, "%d files to do, in %d groups\n", end_file-start_file,
n_file_groups);
new_temp_filename = (char *) rr_malloc(300*sizeof(char));
/*
* These n_file_groups sets of files will be done in groups of
* max_files batches each, as temp files numbered
* end_file+1 ... end_file+n_file_groups,
* and then these will be merged into the final result.
*/
for (i = 0; i < n_file_groups; i++) {
/* do files i*max_files through min((i+1)*max_files-1,end_file); */
new_start_file = start_file + (i*max_files);
new_end_file = start_file + ((i+1)*max_files) - 1;
if (new_end_file > end_file) new_end_file = end_file;
sprintf(new_temp_filename,
"%s/%hu%s",
temp_file_root,
end_file+i+1,
temp_file_ext);
new_temp_file = rr_oopen(new_temp_filename);
merge_tempfiles(new_start_file,
new_end_file,
temp_file_root,
temp_file_ext,
max_files,
new_temp_file,
n,
verbosity);
rr_iclose(new_temp_file);
}
merge_tempfiles(end_file+1,
end_file+n_file_groups,
temp_file_root,
temp_file_ext,
max_files,
outfile,
n,
verbosity);
return;
}
/*
* We know we are now doing <= max_files.
*/
temp_file = (FILE **) rr_malloc((end_file+1)*sizeof(FILE *));
temp_filename = (char **) rr_malloc((end_file+1)*sizeof(char *));
for (i=start_file;i<=end_file;i++) {
temp_filename[i] = (char *) rr_malloc(300*sizeof(char));
}
current_ngram = (char **) rr_malloc((end_file+1)*sizeof(char *));
for (i=start_file;i<=end_file;i++) {
current_ngram[i] = (char *) rr_malloc(1000*sizeof(char));
}
current_ngram_count = (int *) rr_malloc((end_file+1)*sizeof(int));
finished = (flag *) rr_malloc(sizeof(flag)*(end_file+1));
/* Open all the temp files for reading */
for (i=start_file;i<=end_file;i++) {
sprintf(temp_filename[i],"%s/%hu%s",
temp_file_root,i,temp_file_ext);
temp_file[i] = rr_iopen(temp_filename[i]);
}
/* Now go through the files simultaneously, and write out the appropriate
ngram counts to the output file. */
for (i=start_file;i<=end_file;i++) {
finished[i] = 0;
if (!rr_feof(temp_file[i])) {
for (j=0;j<=n-1;j++) {
if (fscanf(temp_file[i],"%s",temp_word) != 1) {
if (!rr_feof(temp_file[i]))
quit(-1,"Error reading temp file %s\n",temp_filename[i]);
}else {
if (j==0)
strcpy(current_ngram[i],temp_word);
else {
strcat(current_ngram[i]," ");
strcat(current_ngram[i],temp_word);
}
}
}
if (fscanf(temp_file[i],"%d",¤t_ngram_count[i]) != 1) {
if (!rr_feof(temp_file[i]))
quit(-1,"Error reading temp file %s\n",temp_filename[i]);
}
}
}
all_finished = 0;
while (!all_finished) {
/* Find the smallest current ngram */
strcpy(smallest_ngram,"");
for (i=start_file;i<=end_file;i++) {
if (!finished[i]) {
if (strcmp(smallest_ngram,current_ngram[i]) > 0 ||
(smallest_ngram[0] == '\0'))
strcpy(smallest_ngram,current_ngram[i]);
}
}
/* For each of the files that are currently holding this ngram,
add its count to the temporary count, and read in a new ngram
from the files. */
temp_count = 0;
for (i=start_file;i<=end_file;i++) {
if (!finished[i]) {
if (!strcmp(smallest_ngram,current_ngram[i])) {
temp_count += current_ngram_count[i];
if (!rr_feof(temp_file[i])) {
for (j=0;j<=n-1;j++) {
if (fscanf(temp_file[i],"%s",temp_word) != 1) {
if (!rr_feof(temp_file[i])) {
quit(-1,"Error reading temp file %s\n",temp_filename[i]);
}
}else {
if (j==0)
strcpy(current_ngram[i],temp_word);
else {
strcat(current_ngram[i]," ");
strcat(current_ngram[i],temp_word);
}
}
}
if (fscanf(temp_file[i],"%d",¤t_ngram_count[i]) != 1) {
if (!rr_feof(temp_file[i])) {
quit(-1,"Error reading temp file count %s\n",
temp_filename[i]);
}
}
}
/*
* PWP: Note that the fscanf may have changed the state of
* temp_file[i], so we re-ask the question rather than just
* doing an "else".
*/
if (rr_feof(temp_file[i])) {
finished[i] = 1;
all_finished = 1;
for (j=start_file;j<=end_file;j++) {
if (!finished[j]) {
all_finished = 0;
}
}
}
}
}
}
/*
* PWP: We cannot conditionalize this on (!all_finished) because
* if we do we may have lost the very last count. (Consider the
* case when several files have ran out of data, but the last
* couple have the last count in them.)
*/
if (fprintf(outfile,"%s %d\n",smallest_ngram,temp_count) < 0) {
quit(-1,"Write error encountered while attempting to merge temporary files.\nAborting, but keeping temporary files.\n");
}
}
for (i=start_file;i<=end_file;i++) {
rr_iclose(temp_file[i]);
remove(temp_filename[i]);
}
free(temp_file);
for (i=start_file;i<=end_file;i++) {
free(temp_filename[i]);
}
free(temp_filename);
for (i=start_file;i<=end_file;i++) {
free(current_ngram[i]);
}
free(current_ngram);
free(current_ngram_count);
free(finished);
}
void main (int argc, char **argv) {
ng_t ng;
arpa_lm_t arpa_ng;
char input_string[500];
int num_of_args;
char *args[MAX_ARGS];
char *lm_filename_arpa;
char *lm_filename_binary;
flag told_to_quit;
flag inconsistant_parameters;
flag backoff_from_unk_inc;
flag backoff_from_unk_exc;
flag backoff_from_ccs_inc;
flag backoff_from_ccs_exc;
flag arpa_lm;
flag binary_lm;
flag include_unks;
char *fb_list_filename;
char *probs_stream_filename;
char *annotation_filename;
char *text_stream_filename;
char *oov_filename;
char *ccs_filename;
double log_base;
char wlist_entry[1024];
char current_cc[200];
int current_cc_id;
FILE *context_cues_fp;
int n;
int generate_size = 10000;
int random_seed;
/* Process command line */
report_version(&argc,argv);
if (pc_flagarg(&argc, argv,"-help") ||
argc == 1 ||
(strcmp(argv[1],"-binary") && strcmp(argv[1],"-arpa"))) {
fprintf(stderr,"evallm : Evaluate a language model.\n");
fprintf(stderr,"Usage : evallm [ -binary .binlm | \n");
fprintf(stderr," -arpa .arpa [ -context .ccs ] ]\n");
exit(1);
}
lm_filename_arpa = salloc(pc_stringarg(&argc, argv,"-arpa",""));
if (strcmp(lm_filename_arpa,"")) {
arpa_lm = 1;
}
else {
arpa_lm = 0;
}
lm_filename_binary = salloc(pc_stringarg(&argc, argv,"-binary",""));
if (strcmp(lm_filename_binary,"")) {
binary_lm = 1;
}
else {
binary_lm = 0;
}
if (arpa_lm && binary_lm) {
quit(-1,"Error : Can't use both -arpa and -binary flags.\n");
}
if (!arpa_lm && !binary_lm) {
quit(-1,"Error : Must specify either a binary or an arpa format language model.\n");
}
ccs_filename = salloc(pc_stringarg(&argc, argv,"-context",""));
if (binary_lm && strcmp(ccs_filename,"")) {
fprintf(stderr,"Warning - context cues file not needed with binary language model file.\nWill ignore it.\n");
}
pc_report_unk_args(&argc,argv,2);
/* Load language model */
if (arpa_lm) {
fprintf(stderr,"Reading in language model from file %s\n",
lm_filename_arpa);
load_arpa_lm(&arpa_ng,lm_filename_arpa);
}
else {
fprintf(stderr,"Reading in language model from file %s\n",
lm_filename_binary);
load_lm(&ng,lm_filename_binary);
}
fprintf(stderr,"\nDone.\n");
if (!arpa_lm) {
n=ng.n;
}
else {
n=arpa_ng.n;
}
if (arpa_lm) {
arpa_ng.context_cue =
(flag *) rr_calloc(arpa_ng.table_sizes[0],sizeof(flag));
arpa_ng.no_of_ccs = 0;
if (strcmp(ccs_filename,"")) {
context_cues_fp = rr_iopen(ccs_filename);
while (fgets (wlist_entry, sizeof (wlist_entry),context_cues_fp)) {
if (strncmp(wlist_entry,"##",2)==0) continue;
sscanf (wlist_entry, "%s ",current_cc);
if (strncmp(wlist_entry,"#",1)==0) {
fprintf(stderr,"\n\n===========================================================\n");
fprintf(stderr,":\nWARNING: line assumed NOT a comment:\n");
fprintf(stderr, ">>> %s <<<\n",wlist_entry);
fprintf(stderr, " '%s' will be included in the context cues list\n",current_cc);
fprintf(stderr, " (comments must start with '##')\n");
fprintf(stderr,"===========================================================\n\n");
}
if (sih_lookup(arpa_ng.vocab_ht,current_cc,¤t_cc_id) == 0) {
quit(-1,"Error : %s in the context cues file does not appear in the vocabulary.\n",current_cc);
}
arpa_ng.context_cue[(unsigned short) current_cc_id] = 1;
arpa_ng.no_of_ccs++;
fprintf(stderr,"Context cue word : %s id = %d\n",current_cc,current_cc_id);
}
rr_iclose(context_cues_fp);
}
}
/* Process commands */
told_to_quit = 0;
num_of_args = 0;
while (!feof(stdin) && !told_to_quit) {
printf("evallm : ");
gets(input_string);
if (!feof(stdin)) {
parse_comline(input_string,&num_of_args,args);
random_seed = pc_intarg(&num_of_args,args,"-seed",-1);
generate_size = pc_intarg(&num_of_args,args,"-size",10000);
log_base = pc_doublearg(&num_of_args,args,"-log_base",10.0);
backoff_from_unk_inc = pc_flagarg(&num_of_args,args,
"-backoff_from_unk_inc");
backoff_from_ccs_inc = pc_flagarg(&num_of_args,args,
"-backoff_from_ccs_inc");
backoff_from_unk_exc = pc_flagarg(&num_of_args,args,
"-backoff_from_unk_exc");
backoff_from_ccs_exc = pc_flagarg(&num_of_args,args,
"-backoff_from_ccs_exc");
include_unks = pc_flagarg(&num_of_args,args,"-include_unks");
fb_list_filename = salloc(pc_stringarg(&num_of_args,args,
"-backoff_from_list",""));
text_stream_filename =
salloc(pc_stringarg(&num_of_args,args,"-text",""));
probs_stream_filename =
salloc(pc_stringarg(&num_of_args,args,"-probs",""));
annotation_filename =
salloc(pc_stringarg(&num_of_args,args,"-annotate",""));
oov_filename = salloc(pc_stringarg(&num_of_args,args,"-oovs",""));
inconsistant_parameters = 0;
if (backoff_from_unk_inc && backoff_from_unk_exc) {
fprintf(stderr,"Error : Cannot specify both exclusive and inclusive forced backoff.\n");
fprintf(stderr,"Use only one of -backoff_from_unk_exc and -backoff_from_unk_inc\n");
inconsistant_parameters = 1;
}
if (backoff_from_ccs_inc && backoff_from_ccs_exc) {
fprintf(stderr,"Error : Cannot specify both exclusive and inclusive forced backoff.\n");
fprintf(stderr,"Use only one of -backoff_from_ccs_exc and -backoff_from_ccs_inc\n");
inconsistant_parameters = 1;
}
if (num_of_args > 0) {
if (!inconsistant_parameters) {
if (!strcmp(args[0],"perplexity")) {
compute_perplexity(&ng,
&arpa_ng,
text_stream_filename,
probs_stream_filename,
annotation_filename,
oov_filename,
fb_list_filename,
backoff_from_unk_inc,
backoff_from_unk_exc,
backoff_from_ccs_inc,
backoff_from_ccs_exc,
arpa_lm,
include_unks,
log_base);
}
else {
if (!strcmp(args[0],"validate")) {
if (num_of_args != n) {
fprintf(stderr,"Error : must specify %d words of context.\n",
n-1);
}
else {
/* Assume last n-1 parameters form context */
validate(&ng,
&arpa_ng,
&(args[num_of_args-n+1]),
backoff_from_unk_inc,
backoff_from_unk_exc,
backoff_from_ccs_inc,
backoff_from_ccs_exc,
arpa_lm,
fb_list_filename);
}
}
else {
if (!strcmp(args[0],"stats")) {
if (arpa_lm) {
display_arpa_stats(&arpa_ng);
}
else {
display_stats(&ng);
}
}
else {
if (!strcmp(args[0],"quit")) {
told_to_quit=1;
}
else if (!strcmp(args[0],"generate")) {
if(arpa_lm)
generate_words(NULL,&arpa_ng,generate_size,random_seed,text_stream_filename);
else
generate_words(&ng,NULL,generate_size,random_seed,text_stream_filename);
}
else {
if (!strcmp(args[0],"help")) {
printf("The user may specify one of the following commands: \n");
printf("\n");
printf(" - perplexity\n");
printf("\n");
printf("Computes the perplexity of a given text. May optionally specify words\n");
printf("from which to force back-off.\n");
printf("\n");
printf("Syntax: \n");
printf("\n");
printf("perplexity -text .text\n");
printf(" [ -probs .fprobs ]\n");
printf(" [ -oovs .oov_file ]\n");
printf(" [ -annotate .annotation_file ] \n");
printf(" [ -backoff_from_unk_inc | -backoff_from_unk_exc ]\n");
printf(" [ -backoff_from_ccs_inc | -backoff_from_ccs_exc ] \n");
printf(" [ -backoff_from_list .fblist ]\n");
printf(" [ -include_unks ]\n");
printf("\n");
printf(" - validate\n");
printf(" \n");
printf("Calculate the sum of the probabilities of all the words in the\n");
printf("vocabulary given the context specified by the user.\n");
printf("\n");
printf("Syntax: \n");
printf("\n");
printf("validate [ -backoff_from_unk -backoff_from_ccs |\n");
printf(" -backoff_from_list .fblist ]\n");
printf(" [ -forced_backoff_inc | -forced_back_off_exc ] \n");
printf(" word1 word2 ... word_(n-1)\n");
printf("\n");
printf("Where n is the n in n-gram. \n");
printf("\n");
printf(" - help\n");
printf("\n");
printf("Displays this help message.\n");
printf("\n");
printf("Syntax: \n");
printf("\n");
printf("help\n");
printf("\n");
printf(" - quit\n");
printf("\n");
printf("Exits the program.\n");
printf("\n");
printf("Syntax: \n");
printf("\n");
printf("quit\n");
}
else {
fprintf(stderr,"Unknown command : %s\nType \'help\'\n",
args[0]);
}
}
}
}
}
}
}
}
}
fprintf(stderr,"evallm : Done.\n");
exit(0);
}
int main(int argc, char **argv) {
int i,j;
ng_t* ng;
int verbosity;
int mem_alloc_method; /* Method used to decide how much memory to
allocate for count tables */
int buffer_size;
flag is_ascii;
ngram current_ngram;
ngram previous_ngram;
count_t *ng_count; /* Array indicating the number of occurrances of
the current 1-gram, 2-gram, ... ,n-gram
Size depends on #define in general.h
*/
int nlines;
int pos_of_novelty;
int prev_id1;
flag contains_unks;
int mem_alloced;
flag displayed_oov_warning; /** Display OOV warning
*/
/* ------------------ Process command line --------------------- */
report_version(&argc,argv);
if (argc == 1 || pc_flagarg(&argc, argv,"-help")) {
/* Display help message */
help_message();
exit(1);
}
verbosity = pc_intarg(&argc, argv,"-verbosity",DEFAULT_VERBOSITY);
/* Initialization */
{
ng=init_ng(
&argc,
argv,
verbosity
);
mem_alloc_method = init_alloc_method(ng, &argc, argv, &buffer_size);
if (!strcmp(ng->id_gram_filename,"-") && mem_alloc_method == TWO_PASSES)
quit(-1,"Error: If idngram is read from stdin, then cannot use -calc_mem option.\n");
is_ascii = set_lmformat(pc_flagarg(&argc,argv,"-ascii_input"),
pc_flagarg(&argc,argv,"-bin_input"),
ng);
/* Report parameters */
report_param(verbosity,ng,
is_ascii, mem_alloc_method, buffer_size);
pc_report_unk_args(&argc,argv,verbosity);
}
/* --------------- Read in the vocabulary -------------- */
read_vocab(ng,verbosity);
/* --------------- Allocate space for the table_size array --------- */
init_ng_table_size(ng,
mem_alloc_method,
is_ascii,
verbosity,
buffer_size
);
/* ----------- Allocate memory for tree structure -------------- */
ng->count = NULL;
ng->count4 = NULL;
ng->marg_counts = NULL;
ng->marg_counts4 = NULL;
ng->count_table = NULL;
ng->count = (count_ind_t **) rr_malloc(sizeof(count_ind_t *)*ng->n);
ng->count4 = (count_t **) rr_malloc(sizeof(count_t *)*ng->n);
ng->count_table = (count_t **) rr_malloc(sizeof(count_t *)*ng->n);
if (ng->four_byte_counts) {
ng->marg_counts4 = (count_t *) rr_calloc(sizeof(count_t), ng->table_sizes[0]);
}else {
for (i=0;i<=ng->n-1;i++)
ng->count_table[i] = (count_t *) rr_calloc(ng->count_table_size+1,
sizeof(count_t));
ng->marg_counts = (count_ind_t *) rr_calloc(sizeof(count_ind_t),ng->table_sizes[0]);
fprintf(stderr, "table_size %d\n",ng->table_sizes[0]);
fflush(stderr);
}
ng->word_id = (id__t **) rr_malloc(sizeof(id__t *)*ng->n);
if (ng->four_byte_alphas) {
ng->bo_weight4 = (four_byte_t **) rr_malloc(sizeof(four_byte_t *)*ng->n);
ng->bo_weight4[0] = (four_byte_t *) rr_malloc(sizeof(four_byte_t)*
ng->table_sizes[0]);
}else {
ng->bo_weight = (bo_weight_t **) rr_malloc(sizeof(bo_weight_t *)*ng->n);
ng->bo_weight[0] = (bo_weight_t *) rr_malloc(sizeof(bo_weight_t)*
ng->table_sizes[0]);
}
ng->ind = (index__t **) rr_malloc(sizeof(index__t *)*ng->n);
/* First table */
if (ng->four_byte_counts)
ng->count4[0] = (count_t *) rr_calloc(ng->table_sizes[0],sizeof(count_t));
else
ng->count[0] = (count_ind_t *) rr_calloc(ng->table_sizes[0],sizeof(count_ind_t));
ng->uni_probs = (uni_probs_t *) rr_malloc(sizeof(uni_probs_t)*
ng->table_sizes[0]);
ng->uni_log_probs = (uni_probs_t *) rr_malloc(sizeof(uni_probs_t)*
ng->table_sizes[0]);
if (ng->n >=2)
ng->ind[0] = (index__t *) rr_calloc(ng->table_sizes[0],sizeof(index__t));
for (i=1;i<=ng->n-2;i++) {
ng->word_id[i] = (id__t *) rr_malloc(sizeof(id__t)*ng->table_sizes[i]);
if (ng->four_byte_counts)
ng->count4[i] = (count_t *) rr_malloc(sizeof(count_t)*ng->table_sizes[i]);
else
ng->count[i] = (count_ind_t *) rr_malloc(sizeof(count_ind_t)*ng->table_sizes[i]);
if (ng->four_byte_alphas)
ng->bo_weight4[i] = (four_byte_t *) rr_malloc(sizeof(four_byte_t)*ng->table_sizes[i]);
else
ng->bo_weight[i] = (bo_weight_t *) rr_malloc(sizeof(bo_weight_t)*ng->table_sizes[i]);
ng->ind[i] = (index__t *) rr_malloc(sizeof(index__t)*ng->table_sizes[i]);
mem_alloced = sizeof(count_ind_t) + sizeof(bo_weight_t) +
sizeof(index__t) + sizeof(id__t);
if (ng->four_byte_alphas)
mem_alloced += 4;
mem_alloced *= ng->table_sizes[i];
pc_message(verbosity,2,"Allocated %d bytes to table for %d-grams.\n",
mem_alloced,i+1);
}
ng->word_id[ng->n-1] = (id__t *)
rr_malloc(sizeof(id__t)*ng->table_sizes[ng->n-1]);
if (ng->four_byte_counts)
ng->count4[ng->n-1] = (count_t *) rr_malloc(sizeof(count_t)*ng->table_sizes[ng->n-1]);
else
ng->count[ng->n-1] = (count_ind_t *) rr_malloc(sizeof(count_ind_t)*ng->table_sizes[ng->n-1]);
pc_message(verbosity,2,"Allocated (%d+%d) bytes to table for %d-grams.\n",
ng->four_byte_counts?sizeof(count_t):sizeof(count_ind_t),
sizeof(id__t)*ng->table_sizes[ng->n-1],ng->n);
/* Allocate memory for table for first-byte of indices */
ng_allocate_ptr_table(ng,NULL,0);
/* Allocate memory for alpha array */
ng->alpha_array = (double *) rr_malloc(sizeof(double)*ng->out_of_range_alphas);
ng->size_of_alpha_array = 0;
/* Allocate memory for frequency of frequency information */
ng->freq_of_freq = (fof_t **) rr_malloc(sizeof(fof_t *)*ng->n);
NG_DISC_METH(ng)->allocate_freq_of_freq(ng);
/* Read n-grams into the tree */
pc_message(verbosity,2,"Processing id n-gram file.\n");
pc_message(verbosity,2,"20,000 n-grams processed for each \".\", 1,000,000 for each line.\n");
/* Allocate space for ngrams id arrays */
current_ngram.id_array = (id__t *) rr_calloc(ng->n,sizeof(id__t));
previous_ngram.id_array = (id__t *) rr_calloc(ng->n,sizeof(id__t));
current_ngram.n = ng->n;
previous_ngram.n = ng->n;
ng->num_kgrams = (ngram_sz_t *) rr_calloc(ng->n,sizeof(ngram_sz_t));
ng_count = (count_t *) rr_calloc(ng->n,sizeof(count_t));
nlines = 1;
ng->n_unigrams = 0;
/* Process first n-gram */
get_ngram(ng->id_gram_fp,¤t_ngram,is_ascii);
contains_unks = ngram_chk_contains_unks(¤t_ngram,ng->n);
/* Skip over any unknown words. They will come first, because <UNK>
always has a word ID of zero. */
while (ng->vocab_type == CLOSED_VOCAB && contains_unks){
/* Stop looking if there are no more N-Grams. Of course, this
means training will fail, since there are no unigrams. */
if (get_ngram(ng->id_gram_fp,¤t_ngram,is_ascii) == 0)
break;
contains_unks = ngram_chk_contains_unks(¤t_ngram,ng->n);
}
for (i=0;i<=ng->n-2;i++) {
ng->ind[i][0] = new_index(0,ng->ptr_table[i],&(ng->ptr_table_size[i]),0);
ng->word_id[i+1][0] = current_ngram.id_array[i+1];
ng->num_kgrams[i+1]++;
ng_count[i] = current_ngram.count;
}
ng_count[0] = current_ngram.count;
NG_DISC_METH(ng)->update_freq_of_freq(ng,ng->n-1,current_ngram.count);
store_normal_count(ng,0,current_ngram.count,ng->n-1);
if (current_ngram.count <= ng->cutoffs[ng->n-2])
ng->num_kgrams[ng->n-1]--;
ngram_copy(&previous_ngram,¤t_ngram,ng->n);
prev_id1 = current_ngram.id_array[0];
displayed_oov_warning = 0;
while (!rr_feof(ng->id_gram_fp)) {
if (get_ngram(ng->id_gram_fp,¤t_ngram,is_ascii)) {
if (ng->vocab_type == CLOSED_VOCAB)
contains_unks=ngram_chk_contains_unks(¤t_ngram,ng->n);
if (!contains_unks || ng->vocab_type != CLOSED_VOCAB) {
/* Test for where this ngram differs from last - do we have an
out-of-order ngram? */
pos_of_novelty = ngram_find_pos_of_novelty(¤t_ngram,&previous_ngram,ng->n,nlines);
nlines++;
show_idngram_nlines(nlines, verbosity);
/* Add new n-gram as soon as it is encountered */
/* If all of the positions 2,3,...,n of the n-gram are context
cues then ignore the n-gram. */
if (ng->n > 1) {
NG_DISC_METH(ng)->update_freq_of_freq(ng,ng->n-1,current_ngram.count);
store_normal_count(ng,ng->num_kgrams[ng->n-1],current_ngram.count,ng->n-1);
ng->word_id[ng->n-1][ng->num_kgrams[ng->n-1]] = current_ngram.id_array[ng->n-1];
ng->num_kgrams[ng->n-1]++;
if (ng->num_kgrams[ng->n-1] >= ng->table_sizes[ng->n-1])
quit(-1,"\nMore than %d %d-grams needed to be stored. Rerun with a higher table size.\n",ng->table_sizes[ng->n-1],ng->n);
}
/* Deal with new 2,3,...,(n-1)-grams */
for (i=ng->n-2;i>=MAX(1,pos_of_novelty);i--) {
NG_DISC_METH(ng)->update_freq_of_freq(ng,i,ng_count[i]);
if (ng_count[i] <= ng->cutoffs[i-1])
ng->num_kgrams[i]--;
else
store_normal_count(ng,ng->num_kgrams[i]-1,ng_count[i],i);
ng_count[i] = current_ngram.count;
ng->word_id[i][ng->num_kgrams[i]] = current_ngram.id_array[i];
ng->ind[i][ng->num_kgrams[i]] = new_index(ng->num_kgrams[i+1]-1,
ng->ptr_table[i],
&(ng->ptr_table_size[i]),
ng->num_kgrams[i]);
ng->num_kgrams[i]++;
if (ng->num_kgrams[i] >= ng->table_sizes[i])
quit(-1,"More than %d %d-grams needed to be stored. Rerun with a higher table size.\n",ng->table_sizes[i],i+1);
}
for (i=0;i<=pos_of_novelty-1;i++)
ng_count[i] += current_ngram.count;
/* Deal with new 1-grams */
if (pos_of_novelty == 0) {
if (ng->n>1) {
for (i = prev_id1 + 1; i <= current_ngram.id_array[0]; i++) {
ng->ind[0][i] = new_index(ng->num_kgrams[1]-1,
ng->ptr_table[0],
&(ng->ptr_table_size[0]),
i);
}
prev_id1 = current_ngram.id_array[0];
}
NG_DISC_METH(ng)->update_freq_of_freq(ng,0,ng_count[0]);
if (!ng->context_cue[previous_ngram.id_array[0]]) {
ng->n_unigrams += ng_count[0];
store_normal_count(ng,previous_ngram.id_array[0],ng_count[0],0);
}
store_marginal_count(ng,previous_ngram.id_array[0],ng_count[0],0);
ng_count[0] = current_ngram.count;
}
if (current_ngram.count <= ng->cutoffs[ng->n-2])
ng->num_kgrams[ng->n-1]--;
ngram_copy(&previous_ngram,¤t_ngram,ng->n);
}else {
if (!displayed_oov_warning){
pc_message(verbosity,2,"Warning : id n-gram stream contains OOV's (n-grams will be ignored).\n");
displayed_oov_warning = 1;
}
}
}
}
rr_iclose(ng->id_gram_fp);
for (i=ng->n-2;i>=1;i--) {
NG_DISC_METH(ng)->update_freq_of_freq(ng,i,ng_count[i]);
if (ng_count[i] <= ng->cutoffs[i-1])
ng->num_kgrams[i]--;
else
store_normal_count(ng,ng->num_kgrams[i]-1,ng_count[i],i);
}
NG_DISC_METH(ng)->update_freq_of_freq(ng,0,ng_count[0]);
if (!ng->context_cue[current_ngram.id_array[0]]) {
ng->n_unigrams += ng_count[0];
store_normal_count(ng,current_ngram.id_array[0],ng_count[0],0);
}
store_marginal_count(ng,current_ngram.id_array[0],ng_count[0],0);
if (ng->n>1) {
for (i=current_ngram.id_array[0]+1;i<=ng->vocab_size;i++)
ng->ind[0][i] = new_index(ng->num_kgrams[1],
ng->ptr_table[0],
&(ng->ptr_table_size[0]),
current_ngram.id_array[0]);
}
/* The idngram reading is completed at this point */
pc_message(verbosity,2,"\n");
/* Impose a minimum unigram count, if required */
if (ng->min_unicount > 0) {
int nchanged= 0;
for (i=ng->first_id;i<=ng->vocab_size;i++) {
if ((return_count(ng->four_byte_counts,
ng->count_table[0],
ng->count[0],
ng->count4[0],
i) < ng->min_unicount) && !ng->context_cue[i]) {
/* There was a bug in V2's switch. Look at segment for ABSOLUTE */
NG_DISC_METH(ng)->reduce_ug_freq_of_freq(ng,i);
ng->n_unigrams += (ng->min_unicount - ng->count[0][i]);
store_normal_count(ng,i,ng->min_unicount,0);
nchanged++;
}
}
if (nchanged > 0)
pc_message(verbosity,2,
"Unigram counts of %d words were bumped up to %d.\n",
nchanged,ng->min_unicount);
}
/* Count zeroton information for unigrams */
ng->freq_of_freq[0][0] = 0;
for (i=ng->first_id;i<=ng->vocab_size;i++) {
if (return_count(ng->four_byte_counts,
ng->count_table[0],
ng->count[0],
ng->count4[0],
i) == 0) {
ng->freq_of_freq[0][0]++;
}
}
if (ng->discounting_method == GOOD_TURING) {
for (i=0;i<=ng->n-1;i++)
for (j=1;j<=ng->fof_size[i];j++)
pc_message(verbosity,3,"fof[%d][%d] = %d\n",i,j,ng->freq_of_freq[i][j]);
}
pc_message(verbosity,2,"Calculating discounted counts.\n");
NG_DISC_METH(ng)->compute_discount_aux(ng, verbosity);
/* Smooth unigram distribution, to give some mass to zerotons */
compute_unigram(ng,verbosity);
/* Increment Contexts if using Good-Turing discounting-> No need otherwise,
since all values are discounted anyway. */
if (ng->discounting_method == GOOD_TURING) {
pc_message(verbosity,2,"Incrementing contexts...\n");
for (i=ng->n-1;i>=1;i--)
increment_context(ng,i,verbosity);
}
/* Calculate back-off weights */
pc_message(verbosity,2,"Calculating back-off weights...\n");
for (i=1;i<=ng->n-1;i++)
compute_back_off(ng,i,verbosity);
if (!ng->four_byte_alphas)
pc_message(verbosity,3,"Number of out of range alphas = %d\n",
ng->size_of_alpha_array);
/* Write out LM */
pc_message(verbosity,2,"Writing out language model...\n");
if (ng->write_arpa)
write_arpa_lm(ng,verbosity);
if (ng->write_bin)
write_bin_lm(ng,verbosity);
pc_message(verbosity,0,"idngram2lm : Done.\n");
return 0;
}