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.

313 lines
11 KiB
Python

# Natural Language Toolkit: Interface to BLLIP Parser
#
# Author: David McClosky <dmcc@bigasterisk.com>
#
# Copyright (C) 2001-2020 NLTK Project
# URL: <http://nltk.org/>
# For license information, see LICENSE.TXT
from nltk.parse.api import ParserI
from nltk.tree import Tree
"""
Interface for parsing with BLLIP Parser. Requires the Python
bllipparser module. BllipParser objects can be constructed with the
``BllipParser.from_unified_model_dir`` class method or manually using the
``BllipParser`` constructor. The former is generally easier if you have
a BLLIP Parser unified model directory -- a basic model can be obtained
from NLTK's downloader. More unified parsing models can be obtained with
BLLIP Parser's ModelFetcher (run ``python -m bllipparser.ModelFetcher``
or see docs for ``bllipparser.ModelFetcher.download_and_install_model``).
Basic usage::
# download and install a basic unified parsing model (Wall Street Journal)
# sudo python -m nltk.downloader bllip_wsj_no_aux
>>> from nltk.data import find
>>> model_dir = find('models/bllip_wsj_no_aux').path
>>> bllip = BllipParser.from_unified_model_dir(model_dir)
# 1-best parsing
>>> sentence1 = 'British left waffles on Falklands .'.split()
>>> top_parse = bllip.parse_one(sentence1)
>>> print(top_parse)
(S1
(S
(NP (JJ British) (NN left))
(VP (VBZ waffles) (PP (IN on) (NP (NNP Falklands))))
(. .)))
# n-best parsing
>>> sentence2 = 'Time flies'.split()
>>> all_parses = bllip.parse_all(sentence2)
>>> print(len(all_parses))
50
>>> print(all_parses[0])
(S1 (S (NP (NNP Time)) (VP (VBZ flies))))
# incorporating external tagging constraints (None means unconstrained tag)
>>> constrained1 = bllip.tagged_parse([('Time', 'VB'), ('flies', 'NNS')])
>>> print(next(constrained1))
(S1 (NP (VB Time) (NNS flies)))
>>> constrained2 = bllip.tagged_parse([('Time', 'NN'), ('flies', None)])
>>> print(next(constrained2))
(S1 (NP (NN Time) (VBZ flies)))
References
----------
- Charniak, Eugene. "A maximum-entropy-inspired parser." Proceedings of
the 1st North American chapter of the Association for Computational
Linguistics conference. Association for Computational Linguistics,
2000.
- Charniak, Eugene, and Mark Johnson. "Coarse-to-fine n-best parsing
and MaxEnt discriminative reranking." Proceedings of the 43rd Annual
Meeting on Association for Computational Linguistics. Association
for Computational Linguistics, 2005.
Known issues
------------
Note that BLLIP Parser is not currently threadsafe. Since this module
uses a SWIG interface, it is potentially unsafe to create multiple
``BllipParser`` objects in the same process. BLLIP Parser currently
has issues with non-ASCII text and will raise an error if given any.
See http://pypi.python.org/pypi/bllipparser/ for more information
on BLLIP Parser's Python interface.
"""
__all__ = ["BllipParser"]
# this block allows this module to be imported even if bllipparser isn't
# available
try:
from bllipparser import RerankingParser
from bllipparser.RerankingParser import get_unified_model_parameters
def _ensure_bllip_import_or_error():
pass
except ImportError as ie:
def _ensure_bllip_import_or_error(ie=ie):
raise ImportError("Couldn't import bllipparser module: %s" % ie)
def _ensure_ascii(words):
try:
for i, word in enumerate(words):
word.decode("ascii")
except UnicodeDecodeError:
raise ValueError(
"Token %d (%r) is non-ASCII. BLLIP Parser "
"currently doesn't support non-ASCII inputs." % (i, word)
)
def _scored_parse_to_nltk_tree(scored_parse):
return Tree.fromstring(str(scored_parse.ptb_parse))
class BllipParser(ParserI):
"""
Interface for parsing with BLLIP Parser. BllipParser objects can be
constructed with the ``BllipParser.from_unified_model_dir`` class
method or manually using the ``BllipParser`` constructor.
"""
def __init__(
self,
parser_model=None,
reranker_features=None,
reranker_weights=None,
parser_options=None,
reranker_options=None,
):
"""
Load a BLLIP Parser model from scratch. You'll typically want to
use the ``from_unified_model_dir()`` class method to construct
this object.
:param parser_model: Path to parser model directory
:type parser_model: str
:param reranker_features: Path the reranker model's features file
:type reranker_features: str
:param reranker_weights: Path the reranker model's weights file
:type reranker_weights: str
:param parser_options: optional dictionary of parser options, see
``bllipparser.RerankingParser.RerankingParser.load_parser_options()``
for more information.
:type parser_options: dict(str)
:param reranker_options: optional
dictionary of reranker options, see
``bllipparser.RerankingParser.RerankingParser.load_reranker_model()``
for more information.
:type reranker_options: dict(str)
"""
_ensure_bllip_import_or_error()
parser_options = parser_options or {}
reranker_options = reranker_options or {}
self.rrp = RerankingParser()
self.rrp.load_parser_model(parser_model, **parser_options)
if reranker_features and reranker_weights:
self.rrp.load_reranker_model(
features_filename=reranker_features,
weights_filename=reranker_weights,
**reranker_options
)
def parse(self, sentence):
"""
Use BLLIP Parser to parse a sentence. Takes a sentence as a list
of words; it will be automatically tagged with this BLLIP Parser
instance's tagger.
:return: An iterator that generates parse trees for the sentence
from most likely to least likely.
:param sentence: The sentence to be parsed
:type sentence: list(str)
:rtype: iter(Tree)
"""
_ensure_ascii(sentence)
nbest_list = self.rrp.parse(sentence)
for scored_parse in nbest_list:
yield _scored_parse_to_nltk_tree(scored_parse)
def tagged_parse(self, word_and_tag_pairs):
"""
Use BLLIP to parse a sentence. Takes a sentence as a list of
(word, tag) tuples; the sentence must have already been tokenized
and tagged. BLLIP will attempt to use the tags provided but may
use others if it can't come up with a complete parse subject
to those constraints. You may also specify a tag as ``None``
to leave a token's tag unconstrained.
:return: An iterator that generates parse trees for the sentence
from most likely to least likely.
:param sentence: Input sentence to parse as (word, tag) pairs
:type sentence: list(tuple(str, str))
:rtype: iter(Tree)
"""
words = []
tag_map = {}
for i, (word, tag) in enumerate(word_and_tag_pairs):
words.append(word)
if tag is not None:
tag_map[i] = tag
_ensure_ascii(words)
nbest_list = self.rrp.parse_tagged(words, tag_map)
for scored_parse in nbest_list:
yield _scored_parse_to_nltk_tree(scored_parse)
@classmethod
def from_unified_model_dir(
cls, model_dir, parser_options=None, reranker_options=None
):
"""
Create a ``BllipParser`` object from a unified parsing model
directory. Unified parsing model directories are a standardized
way of storing BLLIP parser and reranker models together on disk.
See ``bllipparser.RerankingParser.get_unified_model_parameters()``
for more information about unified model directories.
:return: A ``BllipParser`` object using the parser and reranker
models in the model directory.
:param model_dir: Path to the unified model directory.
:type model_dir: str
:param parser_options: optional dictionary of parser options, see
``bllipparser.RerankingParser.RerankingParser.load_parser_options()``
for more information.
:type parser_options: dict(str)
:param reranker_options: optional dictionary of reranker options, see
``bllipparser.RerankingParser.RerankingParser.load_reranker_model()``
for more information.
:type reranker_options: dict(str)
:rtype: BllipParser
"""
(
parser_model_dir,
reranker_features_filename,
reranker_weights_filename,
) = get_unified_model_parameters(model_dir)
return cls(
parser_model_dir,
reranker_features_filename,
reranker_weights_filename,
parser_options,
reranker_options,
)
def demo():
"""This assumes the Python module bllipparser is installed."""
# download and install a basic unified parsing model (Wall Street Journal)
# sudo python -m nltk.downloader bllip_wsj_no_aux
from nltk.data import find
model_dir = find("models/bllip_wsj_no_aux").path
print("Loading BLLIP Parsing models...")
# the easiest way to get started is to use a unified model
bllip = BllipParser.from_unified_model_dir(model_dir)
print("Done.")
sentence1 = "British left waffles on Falklands .".split()
sentence2 = "I saw the man with the telescope .".split()
# this sentence is known to fail under the WSJ parsing model
fail1 = "# ! ? : -".split()
for sentence in (sentence1, sentence2, fail1):
print("Sentence: %r" % " ".join(sentence))
try:
tree = next(bllip.parse(sentence))
print(tree)
except StopIteration:
print("(parse failed)")
# n-best parsing demo
for i, parse in enumerate(bllip.parse(sentence1)):
print("parse %d:\n%s" % (i, parse))
# using external POS tag constraints
print(
"forcing 'tree' to be 'NN':",
next(bllip.tagged_parse([("A", None), ("tree", "NN")])),
)
print(
"forcing 'A' to be 'DT' and 'tree' to be 'NNP':",
next(bllip.tagged_parse([("A", "DT"), ("tree", "NNP")])),
)
# constraints don't have to make sense... (though on more complicated
# sentences, they may cause the parse to fail)
print(
"forcing 'A' to be 'NNP':",
next(bllip.tagged_parse([("A", "NNP"), ("tree", None)])),
)
def setup_module(module):
from nose import SkipTest
try:
_ensure_bllip_import_or_error()
except ImportError:
raise SkipTest(
"doctests from nltk.parse.bllip are skipped because "
"the bllipparser module is not installed"
)