/**
* Prune useless disjuncts.
*/
void pp_and_power_prune(Sentence sent, Parse_Options opts)
{
power_table pt;
power_table_alloc(sent, &pt);
power_table_init(sent, &pt);
power_prune(sent, opts, &pt);
if (pp_prune(sent, opts) > 0)
power_prune(sent, opts, &pt);
/* No benefit for now to make additional pp_prune() & power_prune() -
* additional deletions are very rare and even then most of the
* times only one disjunct is deleted. */
power_table_delete(&pt);
return;
}
/** The return value is the number of disjuncts deleted.
* Implementation notes:
* Normally all the identical disjunct-jets are memory shared.
* The suffix_id of each connector serves as its reference count
* in the power table. Each time when a connector that cannot match
* is discovered, its reference count is decreased, and its
* nearest_word field is assigned BAD_WORD. Due to the memory sharing,
* each such an assignment affects immediately all the identical
* disjunct-jets.
* */
static int power_prune(Sentence sent, Parse_Options opts)
{
power_table pt;
prune_context pc;
int N_deleted[2] = {0}; /* [0] counts first deletions, [1] counts dups. */
int total_deleted = 0;
power_table_alloc(sent, &pt);
power_table_init(sent, &pt);
pc.pt = &pt;
pc.power_cost = 0;
pc.null_links = (opts->min_null_count > 0);
pc.N_changed = 1; /* forces it always to make at least two passes */
pc.sent = sent;
while (1)
{
/* left-to-right pass */
for (WordIdx w = 0; w < sent->length; w++)
{
for (Disjunct **dd = &sent->word[w].d; *dd != NULL; /* See: NEXT */)
{
Disjunct *d = *dd; /* just for convenience */
if (d->left == NULL)
{
dd = &d->next; /* NEXT */
continue;
}
bool is_bad = d->left->nearest_word == BAD_WORD;
if (is_bad || left_connector_list_update(&pc, d->left, w, true) < 0)
{
mark_connector_sequence_for_dequeue(d->left, true);
mark_connector_sequence_for_dequeue(d->right, false);
/* discard the current disjunct */
*dd = d->next; /* NEXT - set current disjunct to the next one */
N_deleted[(int)is_bad]++;
continue;
}
dd = &d->next; /* NEXT */
}
clean_table(pt.r_table_size[w], pt.r_table[w]);
}
total_deleted += N_deleted[0] + N_deleted[1];
lgdebug(D_PRUNE, "Debug: l->r pass changed %d and deleted %d (%d+%d)\n",
pc.N_changed, N_deleted[0]+N_deleted[1], N_deleted[0], N_deleted[1]);
if (pc.N_changed == 0 && N_deleted[0] == 0 && N_deleted[1] == 0) break;
pc.N_changed = N_deleted[0] = N_deleted[1] = 0;
/* right-to-left pass */
for (WordIdx w = sent->length-1; w != (WordIdx) -1; w--)
{
for (Disjunct **dd = &sent->word[w].d; *dd != NULL; /* See: NEXT */)
{
Disjunct *d = *dd; /* just for convenience */
if (d->right == NULL)
{
dd = &d->next; /* NEXT */
continue;
}
bool is_bad = d->right->nearest_word == BAD_WORD;
if (is_bad || right_connector_list_update(&pc, d->right, w, true) >= sent->length)
{
mark_connector_sequence_for_dequeue(d->right, true);
mark_connector_sequence_for_dequeue(d->left, false);
/* Discard the current disjunct. */
*dd = d->next; /* NEXT - set current disjunct to the next one */
N_deleted[(int)is_bad]++;
continue;
}
dd = &d->next; /* NEXT */
}
clean_table(pt.l_table_size[w], pt.l_table[w]);
}
total_deleted += N_deleted[0] + N_deleted[1];
lgdebug(D_PRUNE, "Debug: r->l pass changed %d and deleted %d (%d+%d)\n",
pc.N_changed, N_deleted[0]+N_deleted[1], N_deleted[0], N_deleted[1]);
if (pc.N_changed == 0 && N_deleted[0] == 0 && N_deleted[1] == 0) break;
pc.N_changed = N_deleted[0] = N_deleted[1] = 0;
}
power_table_delete(&pt);
lgdebug(D_PRUNE, "Debug: power prune cost: %d\n", pc.power_cost);
print_time(opts, "power pruned");
if (verbosity_level(D_PRUNE))
{
prt_error("\n\\");
prt_error("Debug: After power_pruning:\n\\");
print_disjunct_counts(sent);
}
#ifdef DEBUG
for (WordIdx w = 0; w < sent->length; w++)
{
for (Disjunct *d = sent->word[w].d; NULL != d; d = d->next)
{
for (Connector *c = d->left; NULL != c; c = c->next)
assert(c->nearest_word != BAD_WORD);
for (Connector *c = d->right; NULL != c; c = c->next)
assert(c->nearest_word != BAD_WORD);
}
}
#endif
return total_deleted;
}