/**
* Assumes that the sentence expression lists have been generated.
*/
void prepare_to_parse(Sentence sent, Parse_Options opts)
{
size_t i;
build_sentence_disjuncts(sent, opts->disjunct_cost);
if (verbosity > 2) {
printf("After expanding expressions into disjuncts:");
print_disjunct_counts(sent);
}
print_time(opts, "Built disjuncts");
for (i=0; i<sent->length; i++) {
sent->word[i].d = eliminate_duplicate_disjuncts(sent->word[i].d);
/* Some long Russian sentences can really blow up, here. */
if (resources_exhausted(opts->resources))
return;
}
print_time(opts, "Eliminated duplicate disjuncts");
if (verbosity > 2) {
printf("\nAfter expression pruning and duplicate elimination:\n");
print_disjunct_counts(sent);
}
set_connector_length_limits(sent, opts);
pp_and_power_prune(sent, opts);
}
void my_prepare_to_parse(Sentence sent, Parse_Options opts) {
/* assumes that the sentence expression lists have been generated */
/* this does all the necessary pruning and building of and */
/* structures. */
int i, has_conjunction;
// build_sentence_disjuncts(sent, opts->disjunct_cost);
// if (verbosity > 2) {
//printf("After expanding expressions into disjuncts:") ;
//print_disjunct_counts(sent);
// }
print_time(opts, "Built disjuncts");
for (i=0; i<sent->length; i++) {
sent->word[i].d = eliminate_duplicate_disjuncts(sent->word[i].d);
}
print_time(opts, "Eliminated duplicate disjuncts");
if (verbosity > 2) {
printf("\nAfter expression pruning and duplicate elimination:\n");
print_disjunct_counts(sent);
}
null_links = (opts->min_null_count > 0);
has_conjunction = sentence_contains_conjunction(sent);
set_connector_length_limits(sent, opts);
build_deletable(sent, has_conjunction);
build_effective_dist(sent, has_conjunction);
/* why do we do these here instead of in
first_prepare_to_parse() only? The
reason is that the deletable region
depends on if null links are in use.
with null_links everything is deletable */
if (!has_conjunction) {
pp_and_power_prune(sent, RUTHLESS, opts);
} else {
pp_and_power_prune(sent, GENTLE, opts);
/*if (verbosity > 2) {
printf("\nAfter Gentle power pruning:\n");
print_disjunct_counts(sent);
}
*/
/*print_time(opts, "Finished gentle power pruning"); */
conjunction_prune(sent, opts);
if (verbosity > 2) {
printf("\nAfter conjunction pruning:\n");
print_disjunct_counts(sent);
print_statistics();
}
print_time(opts, "Done conjunction pruning");
build_conjunction_tables(sent);
install_fat_connectors(sent);
install_special_conjunctive_connectors(sent);
if (verbosity > 2) {
printf("After conjunctions, disjuncts counts:\n");
print_disjunct_counts(sent);
}
set_connector_length_limits(sent, opts);
/* have to do this again cause of the
new fat connectors and disjuncts */
print_time(opts, "Constructed fat disjuncts");
prune(sent);
print_time(opts, "Pruned fat disjuncts");
for (i=0; i<sent->length; i++) {
sent->word[i].d = eliminate_duplicate_disjuncts(sent->word[i].d);
}
if (verbosity > 2) {
printf("After pruning and duplicate elimination:\n");
print_disjunct_counts(sent);
}
print_time(opts, "Eliminated duplicate disjuncts (again)");
if (verbosity > 2) print_AND_statistics(sent);
power_prune(sent, RUTHLESS, opts);
}
/*
if (verbosity > 2) {
printf("\nAfter RUTHLESS power-pruning:\n");
print_disjunct_counts(sent);
}
*/
/* print time for power pruning used to be here */
/* now done in power_prune itself */
print_time(opts, "Initialized fast matcher and hash table");
}
/**
* 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);
}
