/**
* We've already built the sentence disjuncts, and we've pruned them
* and power_pruned(GENTLE) them also. The sentence contains a
* conjunction. deletable[][] has been initialized to indicate the
* ranges which may be deleted in the final linkage.
*
* This routine deletes irrelevant disjuncts. It finds them by first
* marking them all as irrelevant, and then marking the ones that
* might be useable. Finally, the unmarked ones are removed.
*/
void conjunction_prune(Sentence sent, Parse_Options opts)
{
Disjunct * d;
int w;
count_context_t *ctxt = sent->count_ctxt;
ctxt->current_resources = opts->resources;
ctxt->deletable = sent->deletable;
count_set_effective_distance(sent);
/* We begin by unmarking all disjuncts. This would not be necessary if
whenever we created a disjunct we cleared its marked field.
I didn't want to search the program for all such places, so
I did this way. XXX FIXME, someday ...
*/
for (w=0; w<sent->length; w++) {
for (d=sent->word[w].d; d != NULL; d=d->next) {
d->marked = FALSE;
}
}
init_fast_matcher(sent);
ctxt->local_sent = sent->word;
ctxt->null_links = (opts->min_null_count > 0);
/*
for (d = sent->word[0].d; d != NULL; d = d->next) {
if ((d->left == NULL) && region_valid(sent, 0, sent->length, d->right, NULL)) {
mark_region(sent, 0, sent->length, d->right, NULL);
d->marked = TRUE;
}
}
mark_region(sent, 0, sent->length, NULL, NULL);
*/
if (ctxt->null_links) {
mark_region(sent, -1, sent->length, NULL, NULL);
} else {
for (w=0; w<sent->length; w++) {
/* consider removing the words [0,w-1] from the beginning
of the sentence */
if (ctxt->deletable[-1][w]) {
for (d = sent->word[w].d; d != NULL; d = d->next) {
if ((d->left == NULL) && region_valid(sent, w, sent->length, d->right, NULL)) {
mark_region(sent, w, sent->length, d->right, NULL);
d->marked = TRUE;
}
}
}
}
}
delete_unmarked_disjuncts(sent);
free_fast_matcher(sent);
ctxt->local_sent = NULL;
ctxt->current_resources = NULL;
ctxt->deletable = NULL;
count_unset_effective_distance(sent);
}
int sentence_parse(Sentence sent, Parse_Options opts) {
int nl;
verbosity = opts->verbosity;
free_sentence_disjuncts(sent);
resources_reset_space(opts->resources);
if (resources_exhausted(opts->resources)) {
sent->num_valid_linkages = 0;
return 0;
}
expression_prune(sent);
print_time(opts, "Finished expression pruning");
prepare_to_parse(sent, opts);
init_fast_matcher(sent);
init_table(sent);
/* A parse set may have been already been built for this sentence,
if it was previously parsed. If so we free it up before building another. */
free_parse_set(sent);
init_x_table(sent);
for (nl = opts->min_null_count;
(nl<=opts->max_null_count) && (!resources_exhausted(opts->resources)); ++nl) {
sent->null_count = nl;
sent->num_linkages_found = parse(sent, sent->null_count, opts);
print_time(opts, "Counted parses");
post_process_linkages(sent, opts);
if (sent->num_valid_linkages > 0) break;
}
free_table(sent);
free_fast_matcher(sent);
print_time(opts, "Finished parse");
return sent->num_valid_linkages;
}
/** Misnamed, this has nothing to do with chart parsing */
static void chart_parse(Sentence sent, Parse_Options opts)
{
int nl;
fast_matcher_t * mchxt;
count_context_t * ctxt;
/* Build lists of disjuncts */
prepare_to_parse(sent, opts);
if (resources_exhausted(opts->resources)) return;
mchxt = alloc_fast_matcher(sent);
ctxt = alloc_count_context(sent->length);
print_time(opts, "Initialized fast matcher");
if (resources_exhausted(opts->resources))
{
free_count_context(ctxt);
free_fast_matcher(mchxt);
return;
}
/* A parse set may have been already been built for this sentence,
* if it was previously parsed. If so we free it up before
* building another. Huh ?? How could that happen? */
free_parse_info(sent->parse_info);
sent->parse_info = parse_info_new(sent->length);
nl = opts->min_null_count;
while (true)
{
Count_bin hist;
s64 total;
if (resources_exhausted(opts->resources)) break;
sent->null_count = nl;
hist = do_parse(sent, mchxt, ctxt, sent->null_count, opts);
total = hist_total(&hist);
if (opts->verbosity > 1)
{
prt_error("Info: Total count with %zu null links: %lld\n",
sent->null_count, total);
}
/* total is 64-bit, num_linkages_found is 32-bit. Clamp */
total = (total > INT_MAX) ? INT_MAX : total;
total = (total < 0) ? INT_MAX : total;
sent->num_linkages_found = (int) total;
print_time(opts, "Counted parses");
select_linkages(sent, mchxt, ctxt, opts);
compute_chosen_disjuncts(sent);
sane_morphism(sent, opts);
post_process_linkages(sent, opts);
if (sent->num_valid_linkages > 0) break;
/* If we are here, then no valid linkages were found.
* If there was a parse overflow, give up now. */
if (PARSE_NUM_OVERFLOW < total) break;
/* loop termination */
if (nl == opts->max_null_count) break;
/* If we are here, we are going round again. Free stuff. */
free_linkages(sent);
nl++;
}
sort_linkages(sent, opts);
free_count_context(ctxt);
free_fast_matcher(mchxt);
}
/**
* classic_parse() -- parse the given sentence.
* Perform parsing, using the original link-grammar parsing algorithm
* given in the original link-grammar papers.
*
* Do the parse with the minimum number of null-links within the range
* specified by opts->min_null_count and opts->max_null_count.
*
* To that end, call do_parse() with an increasing null_count, from
* opts->min_null_count up to (including) opts->max_null_count, until a
* parse is found.
*
* A note about the disjuncts save/restore that is done here:
* To increase the parsing speed, before invoking do_parse(),
* pp_and_power_prune() is invoked to remove connectors which have no
* possibility to connect. It includes a significant optimization when
* null_count==0 that makes a more aggressive removal, but this
* optimization is not appropriate when null_count>0.
*
* So in case this optimization has been done and a complete parse (i.e.
* a parse when null_count==0) is not found, we are left with sentence
* disjuncts which are not appropriate to continue do_parse() tries with
* null_count>0. To solve that, we need to restore the original
* disjuncts of the sentence and call pp_and_power_prune() once again.
*/
void classic_parse(Sentence sent, Parse_Options opts)
{
fast_matcher_t * mchxt = NULL;
count_context_t * ctxt = NULL;
bool pp_and_power_prune_done = false;
Disjunct **disjuncts_copy = NULL;
bool is_null_count_0 = (0 == opts->min_null_count);
int max_null_count = MIN((int)sent->length, opts->max_null_count);
/* Build lists of disjuncts */
prepare_to_parse(sent, opts);
if (resources_exhausted(opts->resources)) return;
if (is_null_count_0 && (0 < max_null_count))
{
/* Save the disjuncts in case we need to parse with null_count>0. */
disjuncts_copy = alloca(sent->length * sizeof(Disjunct *));
for (size_t i = 0; i < sent->length; i++)
disjuncts_copy[i] = disjuncts_dup(sent->word[i].d);
}
for (int nl = opts->min_null_count; nl <= max_null_count; nl++)
{
Count_bin hist;
s64 total;
if (!pp_and_power_prune_done)
{
if (0 != nl)
{
pp_and_power_prune_done = true;
if (is_null_count_0)
opts->min_null_count = 1; /* Don't optimize for null_count==0. */
/* We are parsing now with null_count>0, when previously we
* parsed with null_count==0. Restore the save disjuncts. */
if (NULL != disjuncts_copy)
{
free_sentence_disjuncts(sent);
for (size_t i = 0; i < sent->length; i++)
sent->word[i].d = disjuncts_copy[i];
disjuncts_copy = NULL;
}
}
pp_and_power_prune(sent, opts);
if (is_null_count_0) opts->min_null_count = 0;
if (resources_exhausted(opts->resources)) break;
free_count_context(ctxt, sent);
free_fast_matcher(sent, mchxt);
pack_sentence(sent);
ctxt = alloc_count_context(sent);
mchxt = alloc_fast_matcher(sent);
print_time(opts, "Initialized fast matcher");
}
if (resources_exhausted(opts->resources)) break;
free_linkages(sent);
sent->null_count = nl;
hist = do_parse(sent, mchxt, ctxt, sent->null_count, opts);
total = hist_total(&hist);
lgdebug(D_PARSE, "Info: Total count with %zu null links: %lld\n",
sent->null_count, total);
/* total is 64-bit, num_linkages_found is 32-bit. Clamp */
total = (total > INT_MAX) ? INT_MAX : total;
total = (total < 0) ? INT_MAX : total;
sent->num_linkages_found = (int) total;
print_time(opts, "Counted parses");
extractor_t * pex = extractor_new(sent->length, sent->rand_state);
bool ovfl = setup_linkages(sent, pex, mchxt, ctxt, opts);
process_linkages(sent, pex, ovfl, opts);
free_extractor(pex);
post_process_lkgs(sent, opts);
if (sent->num_valid_linkages > 0) break;
if ((0 == nl) && (0 < max_null_count) && verbosity > 0)
prt_error("No complete linkages found.\n");
/* If we are here, then no valid linkages were found.
* If there was a parse overflow, give up now. */
if (PARSE_NUM_OVERFLOW < total) break;
//if (sent->num_linkages_found > 0 && nl>0) printf("NUM_LINKAGES_FOUND %d\n", sent->num_linkages_found);
}
sort_linkages(sent, opts);
if (NULL != disjuncts_copy)
{
for (size_t i = 0; i < sent->length; i++)
free_disjuncts(disjuncts_copy[i]);
}
free_count_context(ctxt, sent);
free_fast_matcher(sent, mchxt);
}