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.

682 lines
22 KiB
Python

# -*- coding: utf-8 -*-
# Natural Language Toolkit: Chart Parser for Feature-Based Grammars
#
# Copyright (C) 2001-2019 NLTK Project
# Author: Rob Speer <rspeer@mit.edu>
# Peter Ljunglöf <peter.ljunglof@heatherleaf.se>
# URL: <http://nltk.org/>
# For license information, see LICENSE.TXT
"""
Extension of chart parsing implementation to handle grammars with
feature structures as nodes.
"""
from __future__ import print_function, unicode_literals
from six.moves import range
from nltk.compat import python_2_unicode_compatible
from nltk.featstruct import FeatStruct, unify, TYPE, find_variables
from nltk.sem import logic
from nltk.tree import Tree
from nltk.grammar import (
Nonterminal,
Production,
CFG,
FeatStructNonterminal,
is_nonterminal,
is_terminal,
)
from nltk.parse.chart import (
TreeEdge,
Chart,
ChartParser,
EdgeI,
FundamentalRule,
LeafInitRule,
EmptyPredictRule,
BottomUpPredictRule,
SingleEdgeFundamentalRule,
BottomUpPredictCombineRule,
CachedTopDownPredictRule,
TopDownInitRule,
)
# ////////////////////////////////////////////////////////////
# Tree Edge
# ////////////////////////////////////////////////////////////
@python_2_unicode_compatible
class FeatureTreeEdge(TreeEdge):
"""
A specialized tree edge that allows shared variable bindings
between nonterminals on the left-hand side and right-hand side.
Each ``FeatureTreeEdge`` contains a set of ``bindings``, i.e., a
dictionary mapping from variables to values. If the edge is not
complete, then these bindings are simply stored. However, if the
edge is complete, then the constructor applies these bindings to
every nonterminal in the edge whose symbol implements the
interface ``SubstituteBindingsI``.
"""
def __init__(self, span, lhs, rhs, dot=0, bindings=None):
"""
Construct a new edge. If the edge is incomplete (i.e., if
``dot<len(rhs)``), then store the bindings as-is. If the edge
is complete (i.e., if ``dot==len(rhs)``), then apply the
bindings to all nonterminals in ``lhs`` and ``rhs``, and then
clear the bindings. See ``TreeEdge`` for a description of
the other arguments.
"""
if bindings is None:
bindings = {}
# If the edge is complete, then substitute in the bindings,
# and then throw them away. (If we didn't throw them away, we
# might think that 2 complete edges are different just because
# they have different bindings, even though all bindings have
# already been applied.)
if dot == len(rhs) and bindings:
lhs = self._bind(lhs, bindings)
rhs = [self._bind(elt, bindings) for elt in rhs]
bindings = {}
# Initialize the edge.
TreeEdge.__init__(self, span, lhs, rhs, dot)
self._bindings = bindings
self._comparison_key = (self._comparison_key, tuple(sorted(bindings.items())))
@staticmethod
def from_production(production, index):
"""
:return: A new ``TreeEdge`` formed from the given production.
The new edge's left-hand side and right-hand side will
be taken from ``production``; its span will be
``(index,index)``; and its dot position will be ``0``.
:rtype: TreeEdge
"""
return FeatureTreeEdge(
span=(index, index), lhs=production.lhs(), rhs=production.rhs(), dot=0
)
def move_dot_forward(self, new_end, bindings=None):
"""
:return: A new ``FeatureTreeEdge`` formed from this edge.
The new edge's dot position is increased by ``1``,
and its end index will be replaced by ``new_end``.
:rtype: FeatureTreeEdge
:param new_end: The new end index.
:type new_end: int
:param bindings: Bindings for the new edge.
:type bindings: dict
"""
return FeatureTreeEdge(
span=(self._span[0], new_end),
lhs=self._lhs,
rhs=self._rhs,
dot=self._dot + 1,
bindings=bindings,
)
def _bind(self, nt, bindings):
if not isinstance(nt, FeatStructNonterminal):
return nt
return nt.substitute_bindings(bindings)
def next_with_bindings(self):
return self._bind(self.nextsym(), self._bindings)
def bindings(self):
"""
Return a copy of this edge's bindings dictionary.
"""
return self._bindings.copy()
def variables(self):
"""
:return: The set of variables used by this edge.
:rtype: set(Variable)
"""
return find_variables(
[self._lhs]
+ list(self._rhs)
+ list(self._bindings.keys())
+ list(self._bindings.values()),
fs_class=FeatStruct,
)
def __str__(self):
if self.is_complete():
return TreeEdge.__unicode__(self)
else:
bindings = '{%s}' % ', '.join(
'%s: %r' % item for item in sorted(self._bindings.items())
)
return '%s %s' % (TreeEdge.__unicode__(self), bindings)
# ////////////////////////////////////////////////////////////
# A specialized Chart for feature grammars
# ////////////////////////////////////////////////////////////
# TODO: subsumes check when adding new edges
class FeatureChart(Chart):
"""
A Chart for feature grammars.
:see: ``Chart`` for more information.
"""
def select(self, **restrictions):
"""
Returns an iterator over the edges in this chart.
See ``Chart.select`` for more information about the
``restrictions`` on the edges.
"""
# If there are no restrictions, then return all edges.
if restrictions == {}:
return iter(self._edges)
# Find the index corresponding to the given restrictions.
restr_keys = sorted(restrictions.keys())
restr_keys = tuple(restr_keys)
# If it doesn't exist, then create it.
if restr_keys not in self._indexes:
self._add_index(restr_keys)
vals = tuple(
self._get_type_if_possible(restrictions[key]) for key in restr_keys
)
return iter(self._indexes[restr_keys].get(vals, []))
def _add_index(self, restr_keys):
"""
A helper function for ``select``, which creates a new index for
a given set of attributes (aka restriction keys).
"""
# Make sure it's a valid index.
for key in restr_keys:
if not hasattr(EdgeI, key):
raise ValueError('Bad restriction: %s' % key)
# Create the index.
index = self._indexes[restr_keys] = {}
# Add all existing edges to the index.
for edge in self._edges:
vals = tuple(
self._get_type_if_possible(getattr(edge, key)()) for key in restr_keys
)
index.setdefault(vals, []).append(edge)
def _register_with_indexes(self, edge):
"""
A helper function for ``insert``, which registers the new
edge with all existing indexes.
"""
for (restr_keys, index) in self._indexes.items():
vals = tuple(
self._get_type_if_possible(getattr(edge, key)()) for key in restr_keys
)
index.setdefault(vals, []).append(edge)
def _get_type_if_possible(self, item):
"""
Helper function which returns the ``TYPE`` feature of the ``item``,
if it exists, otherwise it returns the ``item`` itself
"""
if isinstance(item, dict) and TYPE in item:
return item[TYPE]
else:
return item
def parses(self, start, tree_class=Tree):
for edge in self.select(start=0, end=self._num_leaves):
if (
(isinstance(edge, FeatureTreeEdge))
and (edge.lhs()[TYPE] == start[TYPE])
and (unify(edge.lhs(), start, rename_vars=True))
):
for tree in self.trees(edge, complete=True, tree_class=tree_class):
yield tree
# ////////////////////////////////////////////////////////////
# Fundamental Rule
# ////////////////////////////////////////////////////////////
class FeatureFundamentalRule(FundamentalRule):
"""
A specialized version of the fundamental rule that operates on
nonterminals whose symbols are ``FeatStructNonterminal``s. Rather
tha simply comparing the nonterminals for equality, they are
unified. Variable bindings from these unifications are collected
and stored in the chart using a ``FeatureTreeEdge``. When a
complete edge is generated, these bindings are applied to all
nonterminals in the edge.
The fundamental rule states that:
- ``[A -> alpha \* B1 beta][i:j]``
- ``[B2 -> gamma \*][j:k]``
licenses the edge:
- ``[A -> alpha B3 \* beta][i:j]``
assuming that B1 and B2 can be unified to generate B3.
"""
def apply(self, chart, grammar, left_edge, right_edge):
# Make sure the rule is applicable.
if not (
left_edge.end() == right_edge.start()
and left_edge.is_incomplete()
and right_edge.is_complete()
and isinstance(left_edge, FeatureTreeEdge)
):
return
found = right_edge.lhs()
nextsym = left_edge.nextsym()
if isinstance(right_edge, FeatureTreeEdge):
if not is_nonterminal(nextsym):
return
if left_edge.nextsym()[TYPE] != right_edge.lhs()[TYPE]:
return
# Create a copy of the bindings.
bindings = left_edge.bindings()
# We rename vars here, because we don't want variables
# from the two different productions to match.
found = found.rename_variables(used_vars=left_edge.variables())
# Unify B1 (left_edge.nextsym) with B2 (right_edge.lhs) to
# generate B3 (result).
result = unify(nextsym, found, bindings, rename_vars=False)
if result is None:
return
else:
if nextsym != found:
return
# Create a copy of the bindings.
bindings = left_edge.bindings()
# Construct the new edge.
new_edge = left_edge.move_dot_forward(right_edge.end(), bindings)
# Add it to the chart, with appropriate child pointers.
if chart.insert_with_backpointer(new_edge, left_edge, right_edge):
yield new_edge
class FeatureSingleEdgeFundamentalRule(SingleEdgeFundamentalRule):
"""
A specialized version of the completer / single edge fundamental rule
that operates on nonterminals whose symbols are ``FeatStructNonterminal``s.
Rather than simply comparing the nonterminals for equality, they are
unified.
"""
_fundamental_rule = FeatureFundamentalRule()
def _apply_complete(self, chart, grammar, right_edge):
fr = self._fundamental_rule
for left_edge in chart.select(
end=right_edge.start(), is_complete=False, nextsym=right_edge.lhs()
):
for new_edge in fr.apply(chart, grammar, left_edge, right_edge):
yield new_edge
def _apply_incomplete(self, chart, grammar, left_edge):
fr = self._fundamental_rule
for right_edge in chart.select(
start=left_edge.end(), is_complete=True, lhs=left_edge.nextsym()
):
for new_edge in fr.apply(chart, grammar, left_edge, right_edge):
yield new_edge
# ////////////////////////////////////////////////////////////
# Top-Down Prediction
# ////////////////////////////////////////////////////////////
class FeatureTopDownInitRule(TopDownInitRule):
def apply(self, chart, grammar):
for prod in grammar.productions(lhs=grammar.start()):
new_edge = FeatureTreeEdge.from_production(prod, 0)
if chart.insert(new_edge, ()):
yield new_edge
class FeatureTopDownPredictRule(CachedTopDownPredictRule):
"""
A specialized version of the (cached) top down predict rule that operates
on nonterminals whose symbols are ``FeatStructNonterminal``s. Rather
than simply comparing the nonterminals for equality, they are
unified.
The top down expand rule states that:
- ``[A -> alpha \* B1 beta][i:j]``
licenses the edge:
- ``[B2 -> \* gamma][j:j]``
for each grammar production ``B2 -> gamma``, assuming that B1
and B2 can be unified.
"""
def apply(self, chart, grammar, edge):
if edge.is_complete():
return
nextsym, index = edge.nextsym(), edge.end()
if not is_nonterminal(nextsym):
return
# If we've already applied this rule to an edge with the same
# next & end, and the chart & grammar have not changed, then
# just return (no new edges to add).
nextsym_with_bindings = edge.next_with_bindings()
done = self._done.get((nextsym_with_bindings, index), (None, None))
if done[0] is chart and done[1] is grammar:
return
for prod in grammar.productions(lhs=nextsym):
# If the left corner in the predicted production is
# leaf, it must match with the input.
if prod.rhs():
first = prod.rhs()[0]
if is_terminal(first):
if index >= chart.num_leaves():
continue
if first != chart.leaf(index):
continue
# We rename vars here, because we don't want variables
# from the two different productions to match.
if unify(prod.lhs(), nextsym_with_bindings, rename_vars=True):
new_edge = FeatureTreeEdge.from_production(prod, edge.end())
if chart.insert(new_edge, ()):
yield new_edge
# Record the fact that we've applied this rule.
self._done[nextsym_with_bindings, index] = (chart, grammar)
# ////////////////////////////////////////////////////////////
# Bottom-Up Prediction
# ////////////////////////////////////////////////////////////
class FeatureBottomUpPredictRule(BottomUpPredictRule):
def apply(self, chart, grammar, edge):
if edge.is_incomplete():
return
for prod in grammar.productions(rhs=edge.lhs()):
if isinstance(edge, FeatureTreeEdge):
_next = prod.rhs()[0]
if not is_nonterminal(_next):
continue
new_edge = FeatureTreeEdge.from_production(prod, edge.start())
if chart.insert(new_edge, ()):
yield new_edge
class FeatureBottomUpPredictCombineRule(BottomUpPredictCombineRule):
def apply(self, chart, grammar, edge):
if edge.is_incomplete():
return
found = edge.lhs()
for prod in grammar.productions(rhs=found):
bindings = {}
if isinstance(edge, FeatureTreeEdge):
_next = prod.rhs()[0]
if not is_nonterminal(_next):
continue
# We rename vars here, because we don't want variables
# from the two different productions to match.
used_vars = find_variables(
(prod.lhs(),) + prod.rhs(), fs_class=FeatStruct
)
found = found.rename_variables(used_vars=used_vars)
result = unify(_next, found, bindings, rename_vars=False)
if result is None:
continue
new_edge = FeatureTreeEdge.from_production(
prod, edge.start()
).move_dot_forward(edge.end(), bindings)
if chart.insert(new_edge, (edge,)):
yield new_edge
class FeatureEmptyPredictRule(EmptyPredictRule):
def apply(self, chart, grammar):
for prod in grammar.productions(empty=True):
for index in range(chart.num_leaves() + 1):
new_edge = FeatureTreeEdge.from_production(prod, index)
if chart.insert(new_edge, ()):
yield new_edge
# ////////////////////////////////////////////////////////////
# Feature Chart Parser
# ////////////////////////////////////////////////////////////
TD_FEATURE_STRATEGY = [
LeafInitRule(),
FeatureTopDownInitRule(),
FeatureTopDownPredictRule(),
FeatureSingleEdgeFundamentalRule(),
]
BU_FEATURE_STRATEGY = [
LeafInitRule(),
FeatureEmptyPredictRule(),
FeatureBottomUpPredictRule(),
FeatureSingleEdgeFundamentalRule(),
]
BU_LC_FEATURE_STRATEGY = [
LeafInitRule(),
FeatureEmptyPredictRule(),
FeatureBottomUpPredictCombineRule(),
FeatureSingleEdgeFundamentalRule(),
]
class FeatureChartParser(ChartParser):
def __init__(
self,
grammar,
strategy=BU_LC_FEATURE_STRATEGY,
trace_chart_width=20,
chart_class=FeatureChart,
**parser_args
):
ChartParser.__init__(
self,
grammar,
strategy=strategy,
trace_chart_width=trace_chart_width,
chart_class=chart_class,
**parser_args
)
class FeatureTopDownChartParser(FeatureChartParser):
def __init__(self, grammar, **parser_args):
FeatureChartParser.__init__(self, grammar, TD_FEATURE_STRATEGY, **parser_args)
class FeatureBottomUpChartParser(FeatureChartParser):
def __init__(self, grammar, **parser_args):
FeatureChartParser.__init__(self, grammar, BU_FEATURE_STRATEGY, **parser_args)
class FeatureBottomUpLeftCornerChartParser(FeatureChartParser):
def __init__(self, grammar, **parser_args):
FeatureChartParser.__init__(
self, grammar, BU_LC_FEATURE_STRATEGY, **parser_args
)
# ////////////////////////////////////////////////////////////
# Instantiate Variable Chart
# ////////////////////////////////////////////////////////////
class InstantiateVarsChart(FeatureChart):
"""
A specialized chart that 'instantiates' variables whose names
start with '@', by replacing them with unique new variables.
In particular, whenever a complete edge is added to the chart, any
variables in the edge's ``lhs`` whose names start with '@' will be
replaced by unique new ``Variable``s.
"""
def __init__(self, tokens):
FeatureChart.__init__(self, tokens)
def initialize(self):
self._instantiated = set()
FeatureChart.initialize(self)
def insert(self, edge, child_pointer_list):
if edge in self._instantiated:
return False
self.instantiate_edge(edge)
return FeatureChart.insert(self, edge, child_pointer_list)
def instantiate_edge(self, edge):
"""
If the edge is a ``FeatureTreeEdge``, and it is complete,
then instantiate all variables whose names start with '@',
by replacing them with unique new variables.
Note that instantiation is done in-place, since the
parsing algorithms might already hold a reference to
the edge for future use.
"""
# If the edge is a leaf, or is not complete, or is
# already in the chart, then just return it as-is.
if not isinstance(edge, FeatureTreeEdge):
return
if not edge.is_complete():
return
if edge in self._edge_to_cpls:
return
# Get a list of variables that need to be instantiated.
# If there are none, then return as-is.
inst_vars = self.inst_vars(edge)
if not inst_vars:
return
# Instantiate the edge!
self._instantiated.add(edge)
edge._lhs = edge.lhs().substitute_bindings(inst_vars)
def inst_vars(self, edge):
return dict(
(var, logic.unique_variable())
for var in edge.lhs().variables()
if var.name.startswith('@')
)
# ////////////////////////////////////////////////////////////
# Demo
# ////////////////////////////////////////////////////////////
def demo_grammar():
from nltk.grammar import FeatureGrammar
return FeatureGrammar.fromstring(
"""
S -> NP VP
PP -> Prep NP
NP -> NP PP
VP -> VP PP
VP -> Verb NP
VP -> Verb
NP -> Det[pl=?x] Noun[pl=?x]
NP -> "John"
NP -> "I"
Det -> "the"
Det -> "my"
Det[-pl] -> "a"
Noun[-pl] -> "dog"
Noun[-pl] -> "cookie"
Verb -> "ate"
Verb -> "saw"
Prep -> "with"
Prep -> "under"
"""
)
def demo(
print_times=True,
print_grammar=True,
print_trees=True,
print_sentence=True,
trace=1,
parser=FeatureChartParser,
sent='I saw John with a dog with my cookie',
):
import sys, time
print()
grammar = demo_grammar()
if print_grammar:
print(grammar)
print()
print("*", parser.__name__)
if print_sentence:
print("Sentence:", sent)
tokens = sent.split()
t = time.clock()
cp = parser(grammar, trace=trace)
chart = cp.chart_parse(tokens)
trees = list(chart.parses(grammar.start()))
if print_times:
print("Time: %s" % (time.clock() - t))
if print_trees:
for tree in trees:
print(tree)
else:
print("Nr trees:", len(trees))
def run_profile():
import profile
profile.run('for i in range(1): demo()', '/tmp/profile.out')
import pstats
p = pstats.Stats('/tmp/profile.out')
p.strip_dirs().sort_stats('time', 'cum').print_stats(60)
p.strip_dirs().sort_stats('cum', 'time').print_stats(60)
if __name__ == '__main__':
from nltk.data import load
demo()
print()
grammar = load('grammars/book_grammars/feat0.fcfg')
cp = FeatureChartParser(grammar, trace=2)
sent = 'Kim likes children'
tokens = sent.split()
trees = cp.parse(tokens)
for tree in trees:
print(tree)