def _apply(rule_a, rule_b):
"""Applies rule_b to rule_a, returning set of derived rules."""
outputs = []
if rule_a.arity == 2:
new_arity = 1 + rule_b.arity
if new_arity <= 2:
# Cannot apply an arity 2 rule to an arity 2 rule because this would lead
# to a rule with 3 different non-terminal indexes, which is disallowed
# by our QCFG conventions.
source_0 = _substitute(rule_a.source, rule_b.source)
target_0 = _substitute(rule_a.target, rule_b.target)
outputs.append((source_0, target_0, new_arity))
# Rule can potentially be applied to either non-terminal in rule_a.
source_1 = _substitute(rule_a.source,
rule_b.source,
nt=qcfg_rule.NT_2)
target_1 = _substitute(rule_a.target,
rule_b.target,
nt=qcfg_rule.NT_2)
outputs.append((source_1, target_1, new_arity))
elif rule_a.arity == 1:
new_arity = rule_b.arity
source = _substitute(rule_a.source, rule_b.source)
target = _substitute(rule_a.target, rule_b.target)
outputs.append((source, target, new_arity))
output_rules = set()
for source, target, arity in outputs:
source, target = rule_utils.canonicalize_nts(source, target, arity)
output_rules.add(
qcfg_rule.QCFGRule(tuple(source), tuple(target), arity))
return output_rules
def _get_rules_for_other_examples(induced_rules, other_examples):
"""Add rules for examples outside of sample that cannot be derived."""
new_rules = set()
for source_str, target_str in other_examples:
goal_rule = qcfg_rule.QCFGRule(tuple(source_str.split()),
tuple(target_str.split()),
arity=0)
if not derivation_utils.can_derive(goal_rule, induced_rules, None):
new_rules.add(goal_rule)
print("Added %s rules for examples outside of initial sample." %
len(new_rules))
return new_rules
def get_exact_match_rules(dataset):
"""Return set of rules for terminal sequences in both source and target."""
matches = set()
for source_str, target_str in dataset:
source = source_str.split()
target = target_str.split()
matches.update(_find_exact_matches(source, target))
exact_match_rules = set()
for match in matches:
rule = qcfg_rule.QCFGRule(source=tuple(match), target=tuple(match), arity=0)
exact_match_rules.add(rule)
return exact_match_rules
def _make_rule(nts, source, target):
"""Canoncalize NT indexes and return QCFGRule."""
arity = len(nts)
source, target = rule_utils.canonicalize_nts(source, target, arity)
return qcfg_rule.QCFGRule(tuple(source), tuple(target), arity)
def _example_to_rule(source_str, target_str):
"""Convert (source, target) example to a QCFGRule."""
return qcfg_rule.QCFGRule(tuple(source_str.split()),
tuple(target_str.split()),
arity=0)