thallada/nlp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Overhaul syntax gen: dual cfd & signature approach

Browse Source 
Still a WIP. Needs a faster tagging method (e.g. spacy).
This commit is contained in:
Tyler Hallada 2016-11-28 16:07:19 -05:00
parent 3ace25b6e2
commit 82a209c771
1 changed files with 93 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

97
syntax_aware_generate.py
View File

@ -1,7 +1,10 @@
import nltk
import operator
import os
import pickle
import random
import nltk
import re
import codecs
from nltk.tree import Tree
from collections import defaultdict
from tqdm import tqdm
@ -9,23 +12,34 @@ from stat_parser import Parser
syntaxes = defaultdict(set)
SYNTAXES_FILE = 'syntaxes.p'
CFDS_FILE = 'cfds.p'
def tree_hash(self):
return hash(tuple(self.leaves()))
Tree.__hash__ = tree_hash
# NOTE: to me: I need to replace nltk parse and tokenization with spacy because it is much faster and less detailed
# which is actually a plus. The problem is that spacy does not create a syntax tree like nltk does. However, it does
# create a dependency tree, which might be good enough for splitting into chunks that can be swapped out between
# corpora. Shitty bus wifi makes it hard to download spacy data and look up the docs.
def generate():
global syntaxes
parser = Parser()
if not os.path.exists(SYNTAXES_FILE):
sents = nltk.corpus.gutenberg.sents('melville-moby_dick.txt')
sents = sents[0:100]
# sents = nltk.corpus.gutenberg.sents('results.txt')
# NOTE: results.txt is a big file of raw text not included in source control, provide your own corpus.
with codecs.open('results.txt', encoding='utf-8') as corpus:
sents = nltk.sent_tokenize(corpus.read())
sents = [sent for sent in sents if len(sent) < 150][0:1500]
for sent in tqdm(sents):
try:
parsed = parser.parse(' '.join(sent))
parsed = parser.parse(sent)
except TypeError:
pass
syntax_signature(parsed, save=True)
@ -34,13 +48,25 @@ def generate():
else:
with open(SYNTAXES_FILE, 'rb+') as pickle_file:
syntaxes = pickle.load(pickle_file)
if not os.path.exists(CFDS_FILE):
# corpus = nltk.corpus.gutenberg.raw('results.txt')
with codecs.open('results.txt', encoding='utf-8') as corpus:
cfds = [make_cfd(corpus.read(), i, exclude_punctuation=False, case_insensitive=True) for i in range(2, 5)]
with open(CFDS_FILE, 'wb+') as pickle_file:
pickle.dump(cfds, pickle_file)
else:
with open(CFDS_FILE, 'rb+') as pickle_file:
cfds = pickle.load(pickle_file)
sents = nltk.corpus.gutenberg.sents('austen-emma.txt')
sents = [sent for sent in sents if len(sent) < 50]
sent = random.choice(sents)
parsed = parser.parse(' '.join(sent))
print(parsed)
print(' '.join(parsed.leaves()))
replaced_tree = tree_replace(parsed)
print('='*30)
replaced_tree = tree_replace(parsed, cfds, [])
print('=' * 30)
print(' '.join(replaced_tree.leaves()))
print(replaced_tree)
@ -72,12 +98,22 @@ def syntax_signature_recurse(tree, save=False):
raise ValueError('Not a nltk.tree.Tree: {}'.format(tree))
def tree_replace(tree):
def tree_replace(tree, cfds, preceding_children=[]):
condition_search = ' '.join([' '.join(child.leaves()) for child in preceding_children]).lower()
sig = syntax_signature(tree)
if sig in syntaxes:
return random.choice(tuple(syntaxes[sig]))
matching_fragments = tuple(syntaxes[sig])
if len(matching_fragments) > 1 and condition_search:
matching_leaves = [' '.join(frag.leaves()) for frag in matching_fragments]
most_common = get_most_common(condition_search, cfds)
candidates = list(set(matching_leaves).intersection(set(most_common)))
if candidates:
return Tree(tree.label(), [random.choice(candidates)])
# find the first element of get_most_common that is also in this list of matching_leaves
return random.choice(matching_fragments)
else:
children = [tree_replace(child) for child in tree if type(child) is Tree]
children = [tree_replace(child, cfds, preceding_children + tree[0:i])
for i, child in enumerate(tree) if type(child) is Tree]
if not children:
# unable to replace this leaf
return tree
@ -85,5 +121,48 @@ def tree_replace(tree):
return Tree(tree.label(), children)
# TODO: this part should definitely be in a different class or module. I need to be able to resuse this method
# among all of my nlp expirements. See notes in this repo for more detail.
def make_cfd(text, n, cfd=None, exclude_punctuation=True, case_insensitive=True):
if not cfd:
cfd = {}
if exclude_punctuation:
nopunct = re.compile('^\w+$')
sentences = nltk.sent_tokenize(text)
for sent in sentences:
sent = nltk.word_tokenize(sent)
if case_insensitive:
sent = [word.lower() for word in sent]
if exclude_punctuation:
sent = [word for word in sent if nopunct.match(word)]
for i in range(len(sent) - (n - 1)):
condition = ' '.join(sent[i:(i + n) - 1])
sample = sent[(i + n) - 1]
if condition in cfd:
if sample in cfd[condition]:
cfd[condition][sample] += 1
else:
cfd[condition].update({sample: 1})
else:
cfd[condition] = {sample: 1}
return cfd
def get_most_common(search, cfds, most_common=None):
if not most_common:
most_common = list()
words = search.split(' ')
for i in reversed(range(len(cfds))):
n = i + 2
if len(words) >= (n - 1):
query = ' '.join(words[len(words) - (n - 1):])
if query in cfds[i]:
most_common.extend([entry[0] for entry in sorted(cfds[i][query].items(),
key=operator.itemgetter(1),
reverse=True)
if entry[0] not in most_common])
return most_common
if __name__ == '__main__':
generate()