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.

833 lines
29 KiB
Python

# Natural Language Toolkit: Glue Semantics
#
# Author: Dan Garrette <dhgarrette@gmail.com>
#
# Copyright (C) 2001-2020 NLTK Project
# URL: <http://nltk.org/>
# For license information, see LICENSE.TXT
import os
from itertools import chain
import nltk
from nltk.internals import Counter
from nltk.tag import UnigramTagger, BigramTagger, TrigramTagger, RegexpTagger
from nltk.sem.logic import (
Expression,
Variable,
VariableExpression,
LambdaExpression,
AbstractVariableExpression,
)
from nltk.sem import drt
from nltk.sem import linearlogic
SPEC_SEMTYPES = {
"a": "ex_quant",
"an": "ex_quant",
"every": "univ_quant",
"the": "def_art",
"no": "no_quant",
"default": "ex_quant",
}
OPTIONAL_RELATIONSHIPS = ["nmod", "vmod", "punct"]
class GlueFormula(object):
def __init__(self, meaning, glue, indices=None):
if not indices:
indices = set()
if isinstance(meaning, str):
self.meaning = Expression.fromstring(meaning)
elif isinstance(meaning, Expression):
self.meaning = meaning
else:
raise RuntimeError(
"Meaning term neither string or expression: %s, %s"
% (meaning, meaning.__class__)
)
if isinstance(glue, str):
self.glue = linearlogic.LinearLogicParser().parse(glue)
elif isinstance(glue, linearlogic.Expression):
self.glue = glue
else:
raise RuntimeError(
"Glue term neither string or expression: %s, %s"
% (glue, glue.__class__)
)
self.indices = indices
def applyto(self, arg):
""" self = (\\x.(walk x), (subj -o f))
arg = (john , subj)
returns ((walk john), f)
"""
if self.indices & arg.indices: # if the sets are NOT disjoint
raise linearlogic.LinearLogicApplicationException(
"'%s' applied to '%s'. Indices are not disjoint." % (self, arg)
)
else: # if the sets ARE disjoint
return_indices = self.indices | arg.indices
try:
return_glue = linearlogic.ApplicationExpression(
self.glue, arg.glue, arg.indices
)
except linearlogic.LinearLogicApplicationException:
raise linearlogic.LinearLogicApplicationException(
"'%s' applied to '%s'" % (self.simplify(), arg.simplify())
)
arg_meaning_abstracted = arg.meaning
if return_indices:
for dep in self.glue.simplify().antecedent.dependencies[
::-1
]: # if self.glue is (A -o B), dep is in A.dependencies
arg_meaning_abstracted = self.make_LambdaExpression(
Variable("v%s" % dep), arg_meaning_abstracted
)
return_meaning = self.meaning.applyto(arg_meaning_abstracted)
return self.__class__(return_meaning, return_glue, return_indices)
def make_VariableExpression(self, name):
return VariableExpression(name)
def make_LambdaExpression(self, variable, term):
return LambdaExpression(variable, term)
def lambda_abstract(self, other):
assert isinstance(other, GlueFormula)
assert isinstance(other.meaning, AbstractVariableExpression)
return self.__class__(
self.make_LambdaExpression(other.meaning.variable, self.meaning),
linearlogic.ImpExpression(other.glue, self.glue),
)
def compile(self, counter=None):
"""From Iddo Lev's PhD Dissertation p108-109"""
if not counter:
counter = Counter()
(compiled_glue, new_forms) = self.glue.simplify().compile_pos(
counter, self.__class__
)
return new_forms + [
self.__class__(self.meaning, compiled_glue, set([counter.get()]))
]
def simplify(self):
return self.__class__(
self.meaning.simplify(), self.glue.simplify(), self.indices
)
def __eq__(self, other):
return (
self.__class__ == other.__class__
and self.meaning == other.meaning
and self.glue == other.glue
)
def __ne__(self, other):
return not self == other
# sorting for use in doctests which must be deterministic
def __lt__(self, other):
return str(self) < str(other)
def __str__(self):
assert isinstance(self.indices, set)
accum = "%s : %s" % (self.meaning, self.glue)
if self.indices:
accum += " : {" + ", ".join(str(index) for index in self.indices) + "}"
return accum
def __repr__(self):
return "%s" % self
class GlueDict(dict):
def __init__(self, filename, encoding=None):
self.filename = filename
self.file_encoding = encoding
self.read_file()
def read_file(self, empty_first=True):
if empty_first:
self.clear()
try:
contents = nltk.data.load(
self.filename, format="text", encoding=self.file_encoding
)
# TODO: the above can't handle zip files, but this should anyway be fixed in nltk.data.load()
except LookupError as e:
try:
contents = nltk.data.load(
"file:" + self.filename, format="text", encoding=self.file_encoding
)
except LookupError:
raise e
lines = contents.splitlines()
for line in lines: # example: 'n : (\\x.(<word> x), (v-or))'
# lambdacalc -^ linear logic -^
line = line.strip() # remove trailing newline
if not len(line):
continue # skip empty lines
if line[0] == "#":
continue # skip commented out lines
parts = line.split(
" : ", 2
) # ['verb', '(\\x.(<word> x), ( subj -o f ))', '[subj]']
glue_formulas = []
paren_count = 0
tuple_start = 0
tuple_comma = 0
relationships = None
if len(parts) > 1:
for (i, c) in enumerate(parts[1]):
if c == "(":
if paren_count == 0: # if it's the first '(' of a tuple
tuple_start = i + 1 # then save the index
paren_count += 1
elif c == ")":
paren_count -= 1
if paren_count == 0: # if it's the last ')' of a tuple
meaning_term = parts[1][
tuple_start:tuple_comma
] # '\\x.(<word> x)'
glue_term = parts[1][tuple_comma + 1 : i] # '(v-r)'
glue_formulas.append(
[meaning_term, glue_term]
) # add the GlueFormula to the list
elif c == ",":
if (
paren_count == 1
): # if it's a comma separating the parts of the tuple
tuple_comma = i # then save the index
elif c == "#": # skip comments at the ends of lines
if (
paren_count != 0
): # if the line hasn't parsed correctly so far
raise RuntimeError(
"Formula syntax is incorrect for entry " + line
)
break # break to the next line
if len(parts) > 2: # if there is a relationship entry at the end
rel_start = parts[2].index("[") + 1
rel_end = parts[2].index("]")
if rel_start == rel_end:
relationships = frozenset()
else:
relationships = frozenset(
r.strip() for r in parts[2][rel_start:rel_end].split(",")
)
try:
start_inheritance = parts[0].index("(")
end_inheritance = parts[0].index(")")
sem = parts[0][:start_inheritance].strip()
supertype = parts[0][start_inheritance + 1 : end_inheritance]
except:
sem = parts[0].strip()
supertype = None
if sem not in self:
self[sem] = {}
if (
relationships is None
): # if not specified for a specific relationship set
# add all relationship entries for parents
if supertype:
for rels in self[supertype]:
if rels not in self[sem]:
self[sem][rels] = []
glue = self[supertype][rels]
self[sem][rels].extend(glue)
self[sem][rels].extend(
glue_formulas
) # add the glue formulas to every rel entry
else:
if None not in self[sem]:
self[sem][None] = []
self[sem][None].extend(
glue_formulas
) # add the glue formulas to every rel entry
else:
if relationships not in self[sem]:
self[sem][relationships] = []
if supertype:
self[sem][relationships].extend(self[supertype][relationships])
self[sem][relationships].extend(
glue_formulas
) # add the glue entry to the dictionary
def __str__(self):
accum = ""
for pos in self:
str_pos = "%s" % pos
for relset in self[pos]:
i = 1
for gf in self[pos][relset]:
if i == 1:
accum += str_pos + ": "
else:
accum += " " * (len(str_pos) + 2)
accum += "%s" % gf
if relset and i == len(self[pos][relset]):
accum += " : %s" % relset
accum += "\n"
i += 1
return accum
def to_glueformula_list(self, depgraph, node=None, counter=None, verbose=False):
if node is None:
# TODO: should it be depgraph.root? Is this code tested?
top = depgraph.nodes[0]
depList = list(chain(*top["deps"].values()))
root = depgraph.nodes[depList[0]]
return self.to_glueformula_list(depgraph, root, Counter(), verbose)
glueformulas = self.lookup(node, depgraph, counter)
for dep_idx in chain(*node["deps"].values()):
dep = depgraph.nodes[dep_idx]
glueformulas.extend(
self.to_glueformula_list(depgraph, dep, counter, verbose)
)
return glueformulas
def lookup(self, node, depgraph, counter):
semtype_names = self.get_semtypes(node)
semtype = None
for name in semtype_names:
if name in self:
semtype = self[name]
break
if semtype is None:
# raise KeyError, "There is no GlueDict entry for sem type '%s' (for '%s')" % (sem, word)
return []
self.add_missing_dependencies(node, depgraph)
lookup = self._lookup_semtype_option(semtype, node, depgraph)
if not len(lookup):
raise KeyError(
"There is no GlueDict entry for sem type of '%s' "
"with tag '%s', and rel '%s'" % (node["word"], node["tag"], node["rel"])
)
return self.get_glueformulas_from_semtype_entry(
lookup, node["word"], node, depgraph, counter
)
def add_missing_dependencies(self, node, depgraph):
rel = node["rel"].lower()
if rel == "main":
headnode = depgraph.nodes[node["head"]]
subj = self.lookup_unique("subj", headnode, depgraph)
relation = subj["rel"]
node["deps"].setdefault(relation, [])
node["deps"][relation].append(subj["address"])
# node['deps'].append(subj['address'])
def _lookup_semtype_option(self, semtype, node, depgraph):
relationships = frozenset(
depgraph.nodes[dep]["rel"].lower()
for dep in chain(*node["deps"].values())
if depgraph.nodes[dep]["rel"].lower() not in OPTIONAL_RELATIONSHIPS
)
try:
lookup = semtype[relationships]
except KeyError:
# An exact match is not found, so find the best match where
# 'best' is defined as the glue entry whose relationship set has the
# most relations of any possible relationship set that is a subset
# of the actual depgraph
best_match = frozenset()
for relset_option in set(semtype) - set([None]):
if (
len(relset_option) > len(best_match)
and relset_option < relationships
):
best_match = relset_option
if not best_match:
if None in semtype:
best_match = None
else:
return None
lookup = semtype[best_match]
return lookup
def get_semtypes(self, node):
"""
Based on the node, return a list of plausible semtypes in order of
plausibility.
"""
rel = node["rel"].lower()
word = node["word"].lower()
if rel == "spec":
if word in SPEC_SEMTYPES:
return [SPEC_SEMTYPES[word]]
else:
return [SPEC_SEMTYPES["default"]]
elif rel in ["nmod", "vmod"]:
return [node["tag"], rel]
else:
return [node["tag"]]
def get_glueformulas_from_semtype_entry(
self, lookup, word, node, depgraph, counter
):
glueformulas = []
glueFormulaFactory = self.get_GlueFormula_factory()
for meaning, glue in lookup:
gf = glueFormulaFactory(self.get_meaning_formula(meaning, word), glue)
if not len(glueformulas):
gf.word = word
else:
gf.word = "%s%s" % (word, len(glueformulas) + 1)
gf.glue = self.initialize_labels(gf.glue, node, depgraph, counter.get())
glueformulas.append(gf)
return glueformulas
def get_meaning_formula(self, generic, word):
"""
:param generic: A meaning formula string containing the
parameter "<word>"
:param word: The actual word to be replace "<word>"
"""
word = word.replace(".", "")
return generic.replace("<word>", word)
def initialize_labels(self, expr, node, depgraph, unique_index):
if isinstance(expr, linearlogic.AtomicExpression):
name = self.find_label_name(expr.name.lower(), node, depgraph, unique_index)
if name[0].isupper():
return linearlogic.VariableExpression(name)
else:
return linearlogic.ConstantExpression(name)
else:
return linearlogic.ImpExpression(
self.initialize_labels(expr.antecedent, node, depgraph, unique_index),
self.initialize_labels(expr.consequent, node, depgraph, unique_index),
)
def find_label_name(self, name, node, depgraph, unique_index):
try:
dot = name.index(".")
before_dot = name[:dot]
after_dot = name[dot + 1 :]
if before_dot == "super":
return self.find_label_name(
after_dot, depgraph.nodes[node["head"]], depgraph, unique_index
)
else:
return self.find_label_name(
after_dot,
self.lookup_unique(before_dot, node, depgraph),
depgraph,
unique_index,
)
except ValueError:
lbl = self.get_label(node)
if name == "f":
return lbl
elif name == "v":
return "%sv" % lbl
elif name == "r":
return "%sr" % lbl
elif name == "super":
return self.get_label(depgraph.nodes[node["head"]])
elif name == "var":
return "%s%s" % (lbl.upper(), unique_index)
elif name == "a":
return self.get_label(self.lookup_unique("conja", node, depgraph))
elif name == "b":
return self.get_label(self.lookup_unique("conjb", node, depgraph))
else:
return self.get_label(self.lookup_unique(name, node, depgraph))
def get_label(self, node):
"""
Pick an alphabetic character as identifier for an entity in the model.
:param value: where to index into the list of characters
:type value: int
"""
value = node["address"]
letter = [
"f",
"g",
"h",
"i",
"j",
"k",
"l",
"m",
"n",
"o",
"p",
"q",
"r",
"s",
"t",
"u",
"v",
"w",
"x",
"y",
"z",
"a",
"b",
"c",
"d",
"e",
][value - 1]
num = int(value) // 26
if num > 0:
return letter + str(num)
else:
return letter
def lookup_unique(self, rel, node, depgraph):
"""
Lookup 'key'. There should be exactly one item in the associated relation.
"""
deps = [
depgraph.nodes[dep]
for dep in chain(*node["deps"].values())
if depgraph.nodes[dep]["rel"].lower() == rel.lower()
]
if len(deps) == 0:
raise KeyError("'%s' doesn't contain a feature '%s'" % (node["word"], rel))
elif len(deps) > 1:
raise KeyError(
"'%s' should only have one feature '%s'" % (node["word"], rel)
)
else:
return deps[0]
def get_GlueFormula_factory(self):
return GlueFormula
class Glue(object):
def __init__(
self, semtype_file=None, remove_duplicates=False, depparser=None, verbose=False
):
self.verbose = verbose
self.remove_duplicates = remove_duplicates
self.depparser = depparser
from nltk import Prover9
self.prover = Prover9()
if semtype_file:
self.semtype_file = semtype_file
else:
self.semtype_file = os.path.join(
"grammars", "sample_grammars", "glue.semtype"
)
def train_depparser(self, depgraphs=None):
if depgraphs:
self.depparser.train(depgraphs)
else:
self.depparser.train_from_file(
nltk.data.find(
os.path.join("grammars", "sample_grammars", "glue_train.conll")
)
)
def parse_to_meaning(self, sentence):
readings = []
for agenda in self.parse_to_compiled(sentence):
readings.extend(self.get_readings(agenda))
return readings
def get_readings(self, agenda):
readings = []
agenda_length = len(agenda)
atomics = dict()
nonatomics = dict()
while agenda: # is not empty
cur = agenda.pop()
glue_simp = cur.glue.simplify()
if isinstance(
glue_simp, linearlogic.ImpExpression
): # if cur.glue is non-atomic
for key in atomics:
try:
if isinstance(cur.glue, linearlogic.ApplicationExpression):
bindings = cur.glue.bindings
else:
bindings = linearlogic.BindingDict()
glue_simp.antecedent.unify(key, bindings)
for atomic in atomics[key]:
if not (
cur.indices & atomic.indices
): # if the sets of indices are disjoint
try:
agenda.append(cur.applyto(atomic))
except linearlogic.LinearLogicApplicationException:
pass
except linearlogic.UnificationException:
pass
try:
nonatomics[glue_simp.antecedent].append(cur)
except KeyError:
nonatomics[glue_simp.antecedent] = [cur]
else: # else cur.glue is atomic
for key in nonatomics:
for nonatomic in nonatomics[key]:
try:
if isinstance(
nonatomic.glue, linearlogic.ApplicationExpression
):
bindings = nonatomic.glue.bindings
else:
bindings = linearlogic.BindingDict()
glue_simp.unify(key, bindings)
if not (
cur.indices & nonatomic.indices
): # if the sets of indices are disjoint
try:
agenda.append(nonatomic.applyto(cur))
except linearlogic.LinearLogicApplicationException:
pass
except linearlogic.UnificationException:
pass
try:
atomics[glue_simp].append(cur)
except KeyError:
atomics[glue_simp] = [cur]
for entry in atomics:
for gf in atomics[entry]:
if len(gf.indices) == agenda_length:
self._add_to_reading_list(gf, readings)
for entry in nonatomics:
for gf in nonatomics[entry]:
if len(gf.indices) == agenda_length:
self._add_to_reading_list(gf, readings)
return readings
def _add_to_reading_list(self, glueformula, reading_list):
add_reading = True
if self.remove_duplicates:
for reading in reading_list:
try:
if reading.equiv(glueformula.meaning, self.prover):
add_reading = False
break
except Exception as e:
# if there is an exception, the syntax of the formula
# may not be understandable by the prover, so don't
# throw out the reading.
print("Error when checking logical equality of statements", e)
if add_reading:
reading_list.append(glueformula.meaning)
def parse_to_compiled(self, sentence):
gfls = [self.depgraph_to_glue(dg) for dg in self.dep_parse(sentence)]
return [self.gfl_to_compiled(gfl) for gfl in gfls]
def dep_parse(self, sentence):
"""
Return a dependency graph for the sentence.
:param sentence: the sentence to be parsed
:type sentence: list(str)
:rtype: DependencyGraph
"""
# Lazy-initialize the depparser
if self.depparser is None:
from nltk.parse import MaltParser
self.depparser = MaltParser(tagger=self.get_pos_tagger())
if not self.depparser._trained:
self.train_depparser()
return self.depparser.parse(sentence, verbose=self.verbose)
def depgraph_to_glue(self, depgraph):
return self.get_glue_dict().to_glueformula_list(depgraph)
def get_glue_dict(self):
return GlueDict(self.semtype_file)
def gfl_to_compiled(self, gfl):
index_counter = Counter()
return_list = []
for gf in gfl:
return_list.extend(gf.compile(index_counter))
if self.verbose:
print("Compiled Glue Premises:")
for cgf in return_list:
print(cgf)
return return_list
def get_pos_tagger(self):
from nltk.corpus import brown
regexp_tagger = RegexpTagger(
[
(r"^-?[0-9]+(.[0-9]+)?$", "CD"), # cardinal numbers
(r"(The|the|A|a|An|an)$", "AT"), # articles
(r".*able$", "JJ"), # adjectives
(r".*ness$", "NN"), # nouns formed from adjectives
(r".*ly$", "RB"), # adverbs
(r".*s$", "NNS"), # plural nouns
(r".*ing$", "VBG"), # gerunds
(r".*ed$", "VBD"), # past tense verbs
(r".*", "NN"), # nouns (default)
]
)
brown_train = brown.tagged_sents(categories="news")
unigram_tagger = UnigramTagger(brown_train, backoff=regexp_tagger)
bigram_tagger = BigramTagger(brown_train, backoff=unigram_tagger)
trigram_tagger = TrigramTagger(brown_train, backoff=bigram_tagger)
# Override particular words
main_tagger = RegexpTagger(
[(r"(A|a|An|an)$", "ex_quant"), (r"(Every|every|All|all)$", "univ_quant")],
backoff=trigram_tagger,
)
return main_tagger
class DrtGlueFormula(GlueFormula):
def __init__(self, meaning, glue, indices=None):
if not indices:
indices = set()
if isinstance(meaning, str):
self.meaning = drt.DrtExpression.fromstring(meaning)
elif isinstance(meaning, drt.DrtExpression):
self.meaning = meaning
else:
raise RuntimeError(
"Meaning term neither string or expression: %s, %s"
% (meaning, meaning.__class__)
)
if isinstance(glue, str):
self.glue = linearlogic.LinearLogicParser().parse(glue)
elif isinstance(glue, linearlogic.Expression):
self.glue = glue
else:
raise RuntimeError(
"Glue term neither string or expression: %s, %s"
% (glue, glue.__class__)
)
self.indices = indices
def make_VariableExpression(self, name):
return drt.DrtVariableExpression(name)
def make_LambdaExpression(self, variable, term):
return drt.DrtLambdaExpression(variable, term)
class DrtGlueDict(GlueDict):
def get_GlueFormula_factory(self):
return DrtGlueFormula
class DrtGlue(Glue):
def __init__(
self, semtype_file=None, remove_duplicates=False, depparser=None, verbose=False
):
if not semtype_file:
semtype_file = os.path.join(
"grammars", "sample_grammars", "drt_glue.semtype"
)
Glue.__init__(self, semtype_file, remove_duplicates, depparser, verbose)
def get_glue_dict(self):
return DrtGlueDict(self.semtype_file)
def demo(show_example=-1):
from nltk.parse import MaltParser
examples = [
"David sees Mary",
"David eats a sandwich",
"every man chases a dog",
"every man believes a dog sleeps",
"John gives David a sandwich",
"John chases himself",
]
# 'John persuades David to order a pizza',
# 'John tries to go',
# 'John tries to find a unicorn',
# 'John seems to vanish',
# 'a unicorn seems to approach',
# 'every big cat leaves',
# 'every gray cat leaves',
# 'every big gray cat leaves',
# 'a former senator leaves',
print("============== DEMO ==============")
tagger = RegexpTagger(
[
("^(David|Mary|John)$", "NNP"),
(
"^(sees|eats|chases|believes|gives|sleeps|chases|persuades|tries|seems|leaves)$",
"VB",
),
("^(go|order|vanish|find|approach)$", "VB"),
("^(a)$", "ex_quant"),
("^(every)$", "univ_quant"),
("^(sandwich|man|dog|pizza|unicorn|cat|senator)$", "NN"),
("^(big|gray|former)$", "JJ"),
("^(him|himself)$", "PRP"),
]
)
depparser = MaltParser(tagger=tagger)
glue = Glue(depparser=depparser, verbose=False)
for (i, sentence) in enumerate(examples):
if i == show_example or show_example == -1:
print("[[[Example %s]]] %s" % (i, sentence))
for reading in glue.parse_to_meaning(sentence.split()):
print(reading.simplify())
print("")
if __name__ == "__main__":
demo()