""" only replace head nouns """
from nltk.corpus import wordnet as wn
from os.path import join as pjoin
import spacy
import argparse
import os
import random
# load spacy
nlp = spacy.load('en')
def valid(word, replacement):
if replacement.lower() != replacement:
return False
synset_word = wn.synsets(word)
synset_replacement = wn.synsets(replacement)
for item_1 in synset_word:
for item_2 in synset_replacement:
if item_1 == item_2:
return False
# one-step hypernymy/hyponymy check
for item_1 in synset_word:
for subitem in item_1.hypernyms():
for item_2 in synset_replacement:
if item_2 == subitem:
return False
for subitem in item_1.hyponyms():
for item_2 in synset_replacement:
if item_2 == subitem:
return False
return True
def English(word):
for k in range(len(word)):
if word[k] != ' ' and not ('a' <= word[k] <= 'z'):
return False
return True
def tag_noun_chunk(text):
text = text
text_info = nlp(text)
ws = list([x.text for x in text_info])
replace_tag = [False] * len(ws)
for chunk in text_info.noun_chunks:
if not English(chunk.text):
continue
replace_tag[chunk.end - 1] = True
return ws, replace_tag
prefix_list = ['train', 'dev', 'test']
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--data_name', type=str, default='resnet152_nounex_precomp')
parser.add_argument('--data_path', type=str, default='../data')
args = parser.parse_args()
# load all coco captions
coco_captions = list()
for prefix in prefix_list:
os.system('mkdir {:s}/{:s}/{:s}_ex'.format(args.data_path, args.data_name, prefix))
path_coco_captions = pjoin('{:s}/{:s}/'.format(args.data_path, args.data_name), prefix + '_caps.txt')
coco_captions += open(path_coco_captions).readlines()
# count word-frequency for nouns and adjectives
frequency = dict()
for i in range(0, len(coco_captions), 10000):
# default: treat the last word of a noun chunk as its head
caption = ' '.join(coco_captions[i: min(i+10000, len(coco_captions))])
caption = nlp(caption.lower())
for word in caption:
if word.pos_ == 'NOUN' and word.lemma_ != '--PRON--':
word = word.text
if word in frequency:
frequency[word] += 1
else:
frequency[word] = 1
print(i)
# sort and select most frequent nouns
frequency_pair = sorted([(token, frequency[token]) for token in frequency], key=lambda x: x[1], reverse=True)
frequency_pair = list(filter(lambda x: x[1] >= 200, frequency_pair)) # frequency >= 200
# load concreteness
concreteness = list(filter(
lambda x: float(x[1]) > 0.6,
[line.split() for line in open('{:s}/concreteness.txt'.format(args.data_path)).readlines()]))
concrete_words = set([item[0] for item in concreteness])
# filter out concrete words
frequency_pair = list(filter(lambda x: x[0] in concrete_words, frequency_pair))
frequent_words = [item[0] for item in frequency_pair]
frequent_words_set = set(frequent_words)
# get candidates for replacement
candidate_words = dict()
for w in frequent_words:
candidate_words[w] = list(filter(
lambda x: valid(w, x),
[item for item in frequent_words_set]))
# replace these words with words which are similar to themselves but have different meanings
for prefix in prefix_list:
path_coco_captions = pjoin('{:s}/{:s}/'.format(args.data_path, args.data_name), prefix + '_caps.txt')
real_captions = open(path_coco_captions).readlines()
invalid_word_set = set()
for i in range(len(real_captions)):
caption = real_captions[i]
# load 5 recent captions for invalid word set
if i % 5 == 0:
invalid_word_set = set()
for j in range(i, i + 5):
for w in real_captions[j].split():
invalid_word_set.add(w)
replacement_cnt = 0
fout = open('{:s}/{:s}/{:s}_ex/{:d}.txt'.format(args.data_path, args.data_name, prefix, i), 'w')
caption = caption.lower()
words, noun_chunk_tag = tag_noun_chunk(caption)
replaced_words = [w for w in words]
all_replacements = list()
for j, token in enumerate(words):
if not noun_chunk_tag[j]:
continue
if token in candidate_words:
for candidate in candidate_words[token]:
if candidate in invalid_word_set: # skip invalid words
continue
replaced_words[j] = candidate
all_replacements.append(' '.join(replaced_words))
replacement_cnt += 1
replaced_words[j] = words[j]
random.shuffle(all_replacements)
for line in all_replacements:
fout.write(line)
if prefix in ['train', 'dev']:
for j in range(replacement_cnt, 5):
fout.write(' '.join(['<unk>' for _ in caption.strip().split()]) + '\n')
fout.close()
print(prefix, i, replacement_cnt)
