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.

486 lines
14 KiB
Python

# Natural Language Toolkit: Combinatory Categorial Grammar
#
# Copyright (C) 2001-2019 NLTK Project
# Author: Graeme Gange <ggange@csse.unimelb.edu.au>
# URL: <http://nltk.org/>
# For license information, see LICENSE.TXT
"""
The lexicon is constructed by calling
``lexicon.fromstring(<lexicon string>)``.
In order to construct a parser, you also need a rule set.
The standard English rules are provided in chart as
``chart.DefaultRuleSet``.
The parser can then be constructed by calling, for example:
``parser = chart.CCGChartParser(<lexicon>, <ruleset>)``
Parsing is then performed by running
``parser.parse(<sentence>.split())``.
While this returns a list of trees, the default representation
of the produced trees is not very enlightening, particularly
given that it uses the same tree class as the CFG parsers.
It is probably better to call:
``chart.printCCGDerivation(<parse tree extracted from list>)``
which should print a nice representation of the derivation.
This entire process is shown far more clearly in the demonstration:
python chart.py
"""
from __future__ import print_function, division, unicode_literals
import itertools
from six import string_types
from nltk.parse import ParserI
from nltk.parse.chart import AbstractChartRule, EdgeI, Chart
from nltk.tree import Tree
from nltk.ccg.lexicon import fromstring, Token
from nltk.ccg.combinator import (
ForwardT,
BackwardT,
ForwardApplication,
BackwardApplication,
ForwardComposition,
BackwardComposition,
ForwardSubstitution,
BackwardBx,
BackwardSx,
)
from nltk.compat import python_2_unicode_compatible
from nltk.ccg.combinator import *
from nltk.ccg.logic import *
from nltk.sem.logic import *
# Based on the EdgeI class from NLTK.
# A number of the properties of the EdgeI interface don't
# transfer well to CCGs, however.
class CCGEdge(EdgeI):
def __init__(self, span, categ, rule):
self._span = span
self._categ = categ
self._rule = rule
self._comparison_key = (span, categ, rule)
# Accessors
def lhs(self):
return self._categ
def span(self):
return self._span
def start(self):
return self._span[0]
def end(self):
return self._span[1]
def length(self):
return self._span[1] - self.span[0]
def rhs(self):
return ()
def dot(self):
return 0
def is_complete(self):
return True
def is_incomplete(self):
return False
def nextsym(self):
return None
def categ(self):
return self._categ
def rule(self):
return self._rule
class CCGLeafEdge(EdgeI):
'''
Class representing leaf edges in a CCG derivation.
'''
def __init__(self, pos, token, leaf):
self._pos = pos
self._token = token
self._leaf = leaf
self._comparison_key = (pos, token.categ(), leaf)
# Accessors
def lhs(self):
return self._token.categ()
def span(self):
return (self._pos, self._pos + 1)
def start(self):
return self._pos
def end(self):
return self._pos + 1
def length(self):
return 1
def rhs(self):
return self._leaf
def dot(self):
return 0
def is_complete(self):
return True
def is_incomplete(self):
return False
def nextsym(self):
return None
def token(self):
return self._token
def categ(self):
return self._token.categ()
def leaf(self):
return self._leaf
@python_2_unicode_compatible
class BinaryCombinatorRule(AbstractChartRule):
'''
Class implementing application of a binary combinator to a chart.
Takes the directed combinator to apply.
'''
NUMEDGES = 2
def __init__(self, combinator):
self._combinator = combinator
# Apply a combinator
def apply(self, chart, grammar, left_edge, right_edge):
# The left & right edges must be touching.
if not (left_edge.end() == right_edge.start()):
return
# Check if the two edges are permitted to combine.
# If so, generate the corresponding edge.
if self._combinator.can_combine(left_edge.categ(), right_edge.categ()):
for res in self._combinator.combine(left_edge.categ(), right_edge.categ()):
new_edge = CCGEdge(
span=(left_edge.start(), right_edge.end()),
categ=res,
rule=self._combinator,
)
if chart.insert(new_edge, (left_edge, right_edge)):
yield new_edge
# The representation of the combinator (for printing derivations)
def __str__(self):
return "%s" % self._combinator
# Type-raising must be handled slightly differently to the other rules, as the
# resulting rules only span a single edge, rather than both edges.
@python_2_unicode_compatible
class ForwardTypeRaiseRule(AbstractChartRule):
'''
Class for applying forward type raising
'''
NUMEDGES = 2
def __init__(self):
self._combinator = ForwardT
def apply(self, chart, grammar, left_edge, right_edge):
if not (left_edge.end() == right_edge.start()):
return
for res in self._combinator.combine(left_edge.categ(), right_edge.categ()):
new_edge = CCGEdge(span=left_edge.span(), categ=res, rule=self._combinator)
if chart.insert(new_edge, (left_edge,)):
yield new_edge
def __str__(self):
return "%s" % self._combinator
@python_2_unicode_compatible
class BackwardTypeRaiseRule(AbstractChartRule):
'''
Class for applying backward type raising.
'''
NUMEDGES = 2
def __init__(self):
self._combinator = BackwardT
def apply(self, chart, grammar, left_edge, right_edge):
if not (left_edge.end() == right_edge.start()):
return
for res in self._combinator.combine(left_edge.categ(), right_edge.categ()):
new_edge = CCGEdge(span=right_edge.span(), categ=res, rule=self._combinator)
if chart.insert(new_edge, (right_edge,)):
yield new_edge
def __str__(self):
return "%s" % self._combinator
# Common sets of combinators used for English derivations.
ApplicationRuleSet = [
BinaryCombinatorRule(ForwardApplication),
BinaryCombinatorRule(BackwardApplication),
]
CompositionRuleSet = [
BinaryCombinatorRule(ForwardComposition),
BinaryCombinatorRule(BackwardComposition),
BinaryCombinatorRule(BackwardBx),
]
SubstitutionRuleSet = [
BinaryCombinatorRule(ForwardSubstitution),
BinaryCombinatorRule(BackwardSx),
]
TypeRaiseRuleSet = [ForwardTypeRaiseRule(), BackwardTypeRaiseRule()]
# The standard English rule set.
DefaultRuleSet = (
ApplicationRuleSet + CompositionRuleSet + SubstitutionRuleSet + TypeRaiseRuleSet
)
class CCGChartParser(ParserI):
'''
Chart parser for CCGs.
Based largely on the ChartParser class from NLTK.
'''
def __init__(self, lexicon, rules, trace=0):
self._lexicon = lexicon
self._rules = rules
self._trace = trace
def lexicon(self):
return self._lexicon
# Implements the CYK algorithm
def parse(self, tokens):
tokens = list(tokens)
chart = CCGChart(list(tokens))
lex = self._lexicon
# Initialize leaf edges.
for index in range(chart.num_leaves()):
for token in lex.categories(chart.leaf(index)):
new_edge = CCGLeafEdge(index, token, chart.leaf(index))
chart.insert(new_edge, ())
# Select a span for the new edges
for span in range(2, chart.num_leaves() + 1):
for start in range(0, chart.num_leaves() - span + 1):
# Try all possible pairs of edges that could generate
# an edge for that span
for part in range(1, span):
lstart = start
mid = start + part
rend = start + span
for left in chart.select(span=(lstart, mid)):
for right in chart.select(span=(mid, rend)):
# Generate all possible combinations of the two edges
for rule in self._rules:
edges_added_by_rule = 0
for newedge in rule.apply(chart, lex, left, right):
edges_added_by_rule += 1
# Output the resulting parses
return chart.parses(lex.start())
class CCGChart(Chart):
def __init__(self, tokens):
Chart.__init__(self, tokens)
# Constructs the trees for a given parse. Unfortnunately, the parse trees need to be
# constructed slightly differently to those in the default Chart class, so it has to
# be reimplemented
def _trees(self, edge, complete, memo, tree_class):
assert complete, "CCGChart cannot build incomplete trees"
if edge in memo:
return memo[edge]
if isinstance(edge, CCGLeafEdge):
word = tree_class(edge.token(), [self._tokens[edge.start()]])
leaf = tree_class((edge.token(), "Leaf"), [word])
memo[edge] = [leaf]
return [leaf]
memo[edge] = []
trees = []
for cpl in self.child_pointer_lists(edge):
child_choices = [self._trees(cp, complete, memo, tree_class) for cp in cpl]
for children in itertools.product(*child_choices):
lhs = (
Token(
self._tokens[edge.start() : edge.end()],
edge.lhs(),
compute_semantics(children, edge),
),
str(edge.rule()),
)
trees.append(tree_class(lhs, children))
memo[edge] = trees
return trees
def compute_semantics(children, edge):
if children[0].label()[0].semantics() is None:
return None
if len(children) == 2:
if isinstance(edge.rule(), BackwardCombinator):
children = [children[1], children[0]]
combinator = edge.rule()._combinator
function = children[0].label()[0].semantics()
argument = children[1].label()[0].semantics()
if isinstance(combinator, UndirectedFunctionApplication):
return compute_function_semantics(function, argument)
elif isinstance(combinator, UndirectedComposition):
return compute_composition_semantics(function, argument)
elif isinstance(combinator, UndirectedSubstitution):
return compute_substitution_semantics(function, argument)
else:
raise AssertionError('Unsupported combinator \'' + combinator + '\'')
else:
return compute_type_raised_semantics(children[0].label()[0].semantics())
# --------
# Displaying derivations
# --------
def printCCGDerivation(tree):
# Get the leaves and initial categories
leafcats = tree.pos()
leafstr = ''
catstr = ''
# Construct a string with both the leaf word and corresponding
# category aligned.
for (leaf, cat) in leafcats:
str_cat = "%s" % cat
nextlen = 2 + max(len(leaf), len(str_cat))
lcatlen = (nextlen - len(str_cat)) // 2
rcatlen = lcatlen + (nextlen - len(str_cat)) % 2
catstr += ' ' * lcatlen + str_cat + ' ' * rcatlen
lleaflen = (nextlen - len(leaf)) // 2
rleaflen = lleaflen + (nextlen - len(leaf)) % 2
leafstr += ' ' * lleaflen + leaf + ' ' * rleaflen
print(leafstr.rstrip())
print(catstr.rstrip())
# Display the derivation steps
printCCGTree(0, tree)
# Prints the sequence of derivation steps.
def printCCGTree(lwidth, tree):
rwidth = lwidth
# Is a leaf (word).
# Increment the span by the space occupied by the leaf.
if not isinstance(tree, Tree):
return 2 + lwidth + len(tree)
# Find the width of the current derivation step
for child in tree:
rwidth = max(rwidth, printCCGTree(rwidth, child))
# Is a leaf node.
# Don't print anything, but account for the space occupied.
if not isinstance(tree.label(), tuple):
return max(
rwidth, 2 + lwidth + len("%s" % tree.label()), 2 + lwidth + len(tree[0])
)
(token, op) = tree.label()
if op == 'Leaf':
return rwidth
# Pad to the left with spaces, followed by a sequence of '-'
# and the derivation rule.
print(lwidth * ' ' + (rwidth - lwidth) * '-' + "%s" % op)
# Print the resulting category on a new line.
str_res = "%s" % (token.categ())
if token.semantics() is not None:
str_res += " {" + str(token.semantics()) + "}"
respadlen = (rwidth - lwidth - len(str_res)) // 2 + lwidth
print(respadlen * ' ' + str_res)
return rwidth
### Demonstration code
# Construct the lexicon
lex = fromstring(
'''
:- S, NP, N, VP # Primitive categories, S is the target primitive
Det :: NP/N # Family of words
Pro :: NP
TV :: VP/NP
Modal :: (S\\NP)/VP # Backslashes need to be escaped
I => Pro # Word -> Category mapping
you => Pro
the => Det
# Variables have the special keyword 'var'
# '.' prevents permutation
# ',' prevents composition
and => var\\.,var/.,var
which => (N\\N)/(S/NP)
will => Modal # Categories can be either explicit, or families.
might => Modal
cook => TV
eat => TV
mushrooms => N
parsnips => N
bacon => N
'''
)
def demo():
parser = CCGChartParser(lex, DefaultRuleSet)
for parse in parser.parse("I might cook and eat the bacon".split()):
printCCGDerivation(parse)
if __name__ == '__main__':
demo()