You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

341 lines
12 KiB
Python

"""This module implements a CYK parser."""
# Author: https://github.com/ehudt (2018)
#
# Adapted by Erez
from collections import defaultdict
import itertools
from ..exceptions import ParseError
from ..lexer import Token
from ..tree import Tree
from ..grammar import Terminal as T, NonTerminal as NT, Symbol
def match(t, s):
assert isinstance(t, T)
return t.name == s.type
class Rule:
"""Context-free grammar rule."""
def __init__(self, lhs, rhs, weight, alias):
super(Rule, self).__init__()
assert isinstance(lhs, NT), lhs
assert all(isinstance(x, NT) or isinstance(x, T) for x in rhs), rhs
self.lhs = lhs
self.rhs = rhs
self.weight = weight
self.alias = alias
def __str__(self):
return '%s -> %s' % (str(self.lhs), ' '.join(str(x) for x in self.rhs))
def __repr__(self):
return str(self)
def __hash__(self):
return hash((self.lhs, tuple(self.rhs)))
def __eq__(self, other):
return self.lhs == other.lhs and self.rhs == other.rhs
def __ne__(self, other):
return not (self == other)
class Grammar:
"""Context-free grammar."""
def __init__(self, rules):
self.rules = frozenset(rules)
def __eq__(self, other):
return self.rules == other.rules
def __str__(self):
return '\n' + '\n'.join(sorted(repr(x) for x in self.rules)) + '\n'
def __repr__(self):
return str(self)
# Parse tree data structures
class RuleNode:
"""A node in the parse tree, which also contains the full rhs rule."""
def __init__(self, rule, children, weight=0):
self.rule = rule
self.children = children
self.weight = weight
def __repr__(self):
return 'RuleNode(%s, [%s])' % (repr(self.rule.lhs), ', '.join(str(x) for x in self.children))
class Parser:
"""Parser wrapper."""
def __init__(self, rules):
super(Parser, self).__init__()
self.orig_rules = {rule: rule for rule in rules}
rules = [self._to_rule(rule) for rule in rules]
self.grammar = to_cnf(Grammar(rules))
def _to_rule(self, lark_rule):
"""Converts a lark rule, (lhs, rhs, callback, options), to a Rule."""
assert isinstance(lark_rule.origin, NT)
assert all(isinstance(x, Symbol) for x in lark_rule.expansion)
return Rule(
lark_rule.origin, lark_rule.expansion,
weight=lark_rule.options.priority if lark_rule.options.priority else 0,
alias=lark_rule)
def parse(self, tokenized, start): # pylint: disable=invalid-name
"""Parses input, which is a list of tokens."""
assert start
start = NT(start)
table, trees = _parse(tokenized, self.grammar)
# Check if the parse succeeded.
if all(r.lhs != start for r in table[(0, len(tokenized) - 1)]):
raise ParseError('Parsing failed.')
parse = trees[(0, len(tokenized) - 1)][start]
return self._to_tree(revert_cnf(parse))
def _to_tree(self, rule_node):
"""Converts a RuleNode parse tree to a lark Tree."""
orig_rule = self.orig_rules[rule_node.rule.alias]
children = []
for child in rule_node.children:
if isinstance(child, RuleNode):
children.append(self._to_tree(child))
else:
assert isinstance(child.name, Token)
children.append(child.name)
t = Tree(orig_rule.origin, children)
t.rule=orig_rule
return t
def print_parse(node, indent=0):
if isinstance(node, RuleNode):
print(' ' * (indent * 2) + str(node.rule.lhs))
for child in node.children:
print_parse(child, indent + 1)
else:
print(' ' * (indent * 2) + str(node.s))
def _parse(s, g):
"""Parses sentence 's' using CNF grammar 'g'."""
# The CYK table. Indexed with a 2-tuple: (start pos, end pos)
table = defaultdict(set)
# Top-level structure is similar to the CYK table. Each cell is a dict from
# rule name to the best (lightest) tree for that rule.
trees = defaultdict(dict)
# Populate base case with existing terminal production rules
for i, w in enumerate(s):
for terminal, rules in g.terminal_rules.items():
if match(terminal, w):
for rule in rules:
table[(i, i)].add(rule)
if (rule.lhs not in trees[(i, i)] or
rule.weight < trees[(i, i)][rule.lhs].weight):
trees[(i, i)][rule.lhs] = RuleNode(rule, [T(w)], weight=rule.weight)
# Iterate over lengths of sub-sentences
for l in range(2, len(s) + 1):
# Iterate over sub-sentences with the given length
for i in range(len(s) - l + 1):
# Choose partition of the sub-sentence in [1, l)
for p in range(i + 1, i + l):
span1 = (i, p - 1)
span2 = (p, i + l - 1)
for r1, r2 in itertools.product(table[span1], table[span2]):
for rule in g.nonterminal_rules.get((r1.lhs, r2.lhs), []):
table[(i, i + l - 1)].add(rule)
r1_tree = trees[span1][r1.lhs]
r2_tree = trees[span2][r2.lhs]
rule_total_weight = rule.weight + r1_tree.weight + r2_tree.weight
if (rule.lhs not in trees[(i, i + l - 1)]
or rule_total_weight < trees[(i, i + l - 1)][rule.lhs].weight):
trees[(i, i + l - 1)][rule.lhs] = RuleNode(rule, [r1_tree, r2_tree], weight=rule_total_weight)
return table, trees
# This section implements context-free grammar converter to Chomsky normal form.
# It also implements a conversion of parse trees from its CNF to the original
# grammar.
# Overview:
# Applies the following operations in this order:
# * TERM: Eliminates non-solitary terminals from all rules
# * BIN: Eliminates rules with more than 2 symbols on their right-hand-side.
# * UNIT: Eliminates non-terminal unit rules
#
# The following grammar characteristics aren't featured:
# * Start symbol appears on RHS
# * Empty rules (epsilon rules)
class CnfWrapper:
"""CNF wrapper for grammar.
Validates that the input grammar is CNF and provides helper data structures.
"""
def __init__(self, grammar):
super(CnfWrapper, self).__init__()
self.grammar = grammar
self.rules = grammar.rules
self.terminal_rules = defaultdict(list)
self.nonterminal_rules = defaultdict(list)
for r in self.rules:
# Validate that the grammar is CNF and populate auxiliary data structures.
assert isinstance(r.lhs, NT), r
if len(r.rhs) not in [1, 2]:
raise ParseError("CYK doesn't support empty rules")
if len(r.rhs) == 1 and isinstance(r.rhs[0], T):
self.terminal_rules[r.rhs[0]].append(r)
elif len(r.rhs) == 2 and all(isinstance(x, NT) for x in r.rhs):
self.nonterminal_rules[tuple(r.rhs)].append(r)
else:
assert False, r
def __eq__(self, other):
return self.grammar == other.grammar
def __repr__(self):
return repr(self.grammar)
class UnitSkipRule(Rule):
"""A rule that records NTs that were skipped during transformation."""
def __init__(self, lhs, rhs, skipped_rules, weight, alias):
super(UnitSkipRule, self).__init__(lhs, rhs, weight, alias)
self.skipped_rules = skipped_rules
def __eq__(self, other):
return isinstance(other, type(self)) and self.skipped_rules == other.skipped_rules
__hash__ = Rule.__hash__
def build_unit_skiprule(unit_rule, target_rule):
skipped_rules = []
if isinstance(unit_rule, UnitSkipRule):
skipped_rules += unit_rule.skipped_rules
skipped_rules.append(target_rule)
if isinstance(target_rule, UnitSkipRule):
skipped_rules += target_rule.skipped_rules
return UnitSkipRule(unit_rule.lhs, target_rule.rhs, skipped_rules,
weight=unit_rule.weight + target_rule.weight, alias=unit_rule.alias)
def get_any_nt_unit_rule(g):
"""Returns a non-terminal unit rule from 'g', or None if there is none."""
for rule in g.rules:
if len(rule.rhs) == 1 and isinstance(rule.rhs[0], NT):
return rule
return None
def _remove_unit_rule(g, rule):
"""Removes 'rule' from 'g' without changing the language produced by 'g'."""
new_rules = [x for x in g.rules if x != rule]
refs = [x for x in g.rules if x.lhs == rule.rhs[0]]
new_rules += [build_unit_skiprule(rule, ref) for ref in refs]
return Grammar(new_rules)
def _split(rule):
"""Splits a rule whose len(rhs) > 2 into shorter rules."""
rule_str = str(rule.lhs) + '__' + '_'.join(str(x) for x in rule.rhs)
rule_name = '__SP_%s' % (rule_str) + '_%d'
yield Rule(rule.lhs, [rule.rhs[0], NT(rule_name % 1)], weight=rule.weight, alias=rule.alias)
for i in range(1, len(rule.rhs) - 2):
yield Rule(NT(rule_name % i), [rule.rhs[i], NT(rule_name % (i + 1))], weight=0, alias='Split')
yield Rule(NT(rule_name % (len(rule.rhs) - 2)), rule.rhs[-2:], weight=0, alias='Split')
def _term(g):
"""Applies the TERM rule on 'g' (see top comment)."""
all_t = {x for rule in g.rules for x in rule.rhs if isinstance(x, T)}
t_rules = {t: Rule(NT('__T_%s' % str(t)), [t], weight=0, alias='Term') for t in all_t}
new_rules = []
for rule in g.rules:
if len(rule.rhs) > 1 and any(isinstance(x, T) for x in rule.rhs):
new_rhs = [t_rules[x].lhs if isinstance(x, T) else x for x in rule.rhs]
new_rules.append(Rule(rule.lhs, new_rhs, weight=rule.weight, alias=rule.alias))
new_rules.extend(v for k, v in t_rules.items() if k in rule.rhs)
else:
new_rules.append(rule)
return Grammar(new_rules)
def _bin(g):
"""Applies the BIN rule to 'g' (see top comment)."""
new_rules = []
for rule in g.rules:
if len(rule.rhs) > 2:
new_rules += _split(rule)
else:
new_rules.append(rule)
return Grammar(new_rules)
def _unit(g):
"""Applies the UNIT rule to 'g' (see top comment)."""
nt_unit_rule = get_any_nt_unit_rule(g)
while nt_unit_rule:
g = _remove_unit_rule(g, nt_unit_rule)
nt_unit_rule = get_any_nt_unit_rule(g)
return g
def to_cnf(g):
"""Creates a CNF grammar from a general context-free grammar 'g'."""
g = _unit(_bin(_term(g)))
return CnfWrapper(g)
def unroll_unit_skiprule(lhs, orig_rhs, skipped_rules, children, weight, alias):
if not skipped_rules:
return RuleNode(Rule(lhs, orig_rhs, weight=weight, alias=alias), children, weight=weight)
else:
weight = weight - skipped_rules[0].weight
return RuleNode(
Rule(lhs, [skipped_rules[0].lhs], weight=weight, alias=alias), [
unroll_unit_skiprule(skipped_rules[0].lhs, orig_rhs,
skipped_rules[1:], children,
skipped_rules[0].weight, skipped_rules[0].alias)
], weight=weight)
def revert_cnf(node):
"""Reverts a parse tree (RuleNode) to its original non-CNF form (Node)."""
if isinstance(node, T):
return node
# Reverts TERM rule.
if node.rule.lhs.name.startswith('__T_'):
return node.children[0]
else:
children = []
for child in map(revert_cnf, node.children):
# Reverts BIN rule.
if isinstance(child, RuleNode) and child.rule.lhs.name.startswith('__SP_'):
children += child.children
else:
children.append(child)
# Reverts UNIT rule.
if isinstance(node.rule, UnitSkipRule):
return unroll_unit_skiprule(node.rule.lhs, node.rule.rhs,
node.rule.skipped_rules, children,
node.rule.weight, node.rule.alias)
else:
return RuleNode(node.rule, children)