def main(args):
print args
data_dir = args[1]
pdf_dir = os.path.join(data_dir, 'pdfs')
review_dir = os.path.join(data_dir, 'reviews')
if os.path.exists(os.path.join(
data_dir, 'reviews.json')) and not os.path.exists(review_dir):
print 'Loading reviews from a review file'
papers = Paper.from_softconf_dump(
os.path.join(data_dir, 'reviews.json'))
os.makedirs(review_dir)
for paper in papers:
paper.to_json('{}/reviews/{}.json'.format(data_dir, paper.ID))
if not os.path.exists(pdf_dir) or not os.path.exists(review_dir):
print 'PDF/REVIEW dataset must be ready', pdf_dir, review_dir
pdf_files = glob.glob(pdf_dir + '/*.pdf')
print 'Number of pdfs:', len(pdf_files)
review_files = glob.glob(review_dir + '/*.json')
print 'Number of papers:', len(review_files)
# checking the decision distributions
decisions, recs, reviews = [], [], []
category_dict = {}
category_types = ['cs.cl', 'cs.lg', 'cs.ai']
for review_file in review_files:
paper = Paper.from_json(review_file)
reviews += paper.REVIEWS
if not paper: continue
decisions.append(paper.get_accepted())
if len(paper.get_reviews()) > 0:
recs.append(paper.get_reviews()[0].get_recommendation())
# count categories
matched = False
categories = paper.SUBJECTS.lower().split(' ')
for c in categories:
if c in category_types:
matched = c
if matched:
if matched in category_dict:
category_dict[matched] += 1
else:
category_dict[matched] = 0
else:
print categories, paper.ID
print 'Paper Decisions:', Counter(decisions)
print 'Review Recommendations:', Counter(recs)
print 'Number of reviews:', len(reviews)
print 'Categories: ', category_dict
# science parser
print 'Generating science parses...'
science_dir = os.path.join(data_dir, 'scienceparse/')
if not os.path.exists(science_dir):
print 'Parsing papers usnig science-parser...'
os.makedirs(science_dir)
os.system(
'java -Xmx6g -jar ../lib/science-parse-cli-assembly-1.2.9-SNAPSHOT.jar %s -o %s'
% (pdf_dir, science_dir))
science_files = glob.glob(science_dir + '/*.pdf.json')
else:
print 'Reading parsed science parses...'
science_files = glob.glob(science_dir + '/*.pdf.json')
print 'Number of science parses:', len(science_files)
# split to train/dev/test by acceptance
data_types = ['train', 'dev', 'test']
print 'Splitting paper/review/science-parses into train/dev/test'
split_again = False
for dtype in data_types:
data_type_dir = os.path.join(data_dir, dtype)
if os.path.exists(data_type_dir):
num_pdfs = len(glob.glob(data_type_dir + '/pdfs' + '/*.pdf'))
print 'file already exists:', data_type_dir, num_pdfs
else:
split_again = True
if split_again:
print 'splitting ...'
#os.makedirs(data_type_dir)
pids = [
os.path.basename(pfile).replace('.pdf', '') for pfile in pdf_files
]
rids = [
os.path.basename(rfile).replace('.json', '')
for rfile in review_files
]
sids = [
os.path.basename(sfile).replace('.pdf.json', '')
for sfile in science_files
]
ids = []
for pid in pids:
if pid in rids and pid in sids:
ids.append(pid)
train, validtest = train_test_split(ids,
test_size=0.1,
random_state=42)
dev, test = train_test_split(validtest, test_size=0.5, random_state=42)
for didx, data_set in enumerate([train, dev, test]):
if os.path.exists(os.path.join(data_dir, data_types[didx])):
rmtree(os.path.join(data_dir, data_types[didx]))
os.makedirs(os.path.join(data_dir, data_types[didx], 'pdfs'))
os.makedirs(os.path.join(data_dir, data_types[didx], 'reviews'))
os.makedirs(os.path.join(data_dir, data_types[didx],
'parsed_pdfs'))
print 'Splitting..', data_types[didx], len(data_set)
for d in data_set:
copyfile(
os.path.join(pdf_dir, d + '.pdf'),
os.path.join(data_dir, data_types[didx], 'pdfs',
d + '.pdf'))
copyfile(
os.path.join(review_dir, d + '.json'),
os.path.join(data_dir, data_types[didx], 'reviews',
d + '.json'))
copyfile(
os.path.join(science_dir, d + '.pdf.json'),
os.path.join(data_dir, data_types[didx], 'parsed_pdfs',
d + '.pdf.json'))
def main(args, limit=False):
fout_dic = {}
for len_sent in range(args.min_sent, args.max_sent + 1):
fout_dic[len_sent] = open('%s_%d.txt' % (args.out_file, len_sent), 'w')
review_files = glob.glob(args.data_dir + '/reviews' + '/*.json')
print 'Number of papers:', len(review_files)
cnt = 0
topic = ''
topic_changed = False
category_types = ['cs.cl', 'cs.lg', 'cs.ai']
#nlp = spacy.load('en', parser=False)
for rid, review_file in enumerate(review_files):
if rid % 1000 == 0:
print '[%d/%d]' % (rid, len(review_files))
paper = Paper.from_json(review_file)
if not paper:
continue
paper.SCIENCEPARSE = ScienceParseReader.read_science_parse(
paper.ID, paper.TITLE, paper.ABSTRACT, args.data_dir + '/scienceparse/')
# paper ID
file_prefix = paper.ID
# sentences
sections = paper.SCIENCEPARSE.get_sections_dict()
for topic, content in sections.items():
paragraphs = content.split('\n')
for paragraph in paragraphs:
if paragraph == '':
continue
sents = sentence_tokenizer(
paragraph,
min_sent=args.min_sent,
max_sent=args.max_sent,
min_sent_len=args.min_sent_len,
max_sent_len=args.max_sent_len
)
if sents is None:
continue
cnt += 1
avg_len = np.average([len(sent) for sent in sents])
fout = fout_dic[len(sents)]
fout.write('%s\t%s\t%d\t%.2f\t%s\n' % (topic, file_prefix, len(
sents), avg_len, '\t'.join([' '.join(sent) for sent in sents])))
if cnt % 1000 == 0:
print '\t%d paragraphs, %d files ..' % (cnt, rid)
fout.flush()
if limit and cnt == limit:
sys.exit(1)
for len_sent, fout in fout_dic.items():
fout.close()
def prepare_data(data_dir,
vocab_path='vocab',
max_vocab_size=20000,
max_len_paper=1000,
max_len_review=200):
data_type = data_dir.split('/')[-1]
vocab_path += '.' + data_type
if max_vocab_size: vocab_path += '.' + str(max_vocab_size)
vocab_path = data_dir + '/' + vocab_path
label_scale = 5
if 'iclr' in data_dir.lower():
fill_missing = False
aspects = [
'RECOMMENDATION', 'SUBSTANCE', 'APPROPRIATENESS',
'MEANINGFUL_COMPARISON', 'SOUNDNESS_CORRECTNESS', 'ORIGINALITY',
'CLARITY', 'IMPACT', 'RECOMMENDATION_ORIGINAL'
]
review_dir_postfix = '_annotated'
elif 'acl' in data_dir.lower():
fill_missing = True
aspects = [
'RECOMMENDATION', 'SUBSTANCE', 'APPROPRIATENESS',
'MEANINGFUL_COMPARISON', 'SOUNDNESS_CORRECTNESS', 'ORIGINALITY',
'CLARITY', 'IMPACT', 'REVIEWER_CONFIDENCE'
]
review_dir_postfix = ''
else:
print('wrong dataset:', data_dir)
sys.exit(1)
# Loading datasets
print('Reading datasets..')
datasets = ['train', 'dev', 'test']
paper_content_all = []
review_content_all = []
data = defaultdict(list)
for dataset in datasets:
review_dir = os.path.join(data_dir, dataset,
'reviews%s/' % (review_dir_postfix))
scienceparse_dir = os.path.join(data_dir, dataset, 'parsed_pdfs/')
model_dir = os.path.join(data_dir, dataset, 'model/')
if not os.path.exists(model_dir): os.makedirs(model_dir)
paper_json_filenames = sorted(glob.glob(
'{}/*.json'.format(review_dir)))
# add all paper/review content to generate corpus for buildinb vocab
paper_content = []
review_content = []
for paper_json_filename in paper_json_filenames:
d = {}
paper = Paper.from_json(paper_json_filename)
paper.SCIENCEPARSE = ScienceParseReader.read_science_parse(
paper.ID, paper.TITLE, paper.ABSTRACT, scienceparse_dir)
review_contents = []
reviews = []
for review in paper.REVIEWS:
review_contents.append(
preprocess(review.COMMENTS,
only_char=False,
lower=True,
stop_remove=False))
reviews.append(review)
d['paper_content'] = preprocess(
paper.SCIENCEPARSE.get_paper_content(),
only_char=False,
lower=True,
stop_remove=False)
d['reviews_content'] = review_contents
d['reviews'] = reviews
data[dataset].append(d)
print('Total number of papers %d' %
(np.sum([len(d) for _, d in list(data.items())])))
print(
'Total number of reviews %d' %
(np.sum([len(r['reviews']) for _, d in list(data.items())
for r in d])))
# Loading VOCAB
print('Building vocab...')
words = []
for _, d in list(data.items()):
for p in d:
words += p['paper_content'].split(' ')
for r in p['reviews_content']:
words += r.split(' ')
print("Total words in corpus", len(words))
vocab = OrderedDict()
word_counter = Counter(words)
vocab['PAD'] = 0
vocab['UNK'] = 1
for w, c in word_counter.most_common():
if max_vocab_size:
if len(vocab) >= max_vocab_size:
break
if len(w) and w not in vocab:
vocab[w] = len(vocab)
with open(vocab_path, 'w') as fout:
for w, id in list(vocab.items()):
fout.write('%s\t%s\n' % (w, id))
vocab_inv = {int(i): v for v, i in list(vocab.items())}
print("Total vocab of size", len(vocab))
# Loading DATA
print('Reading reviews from...')
data_padded = []
for dataset in datasets:
ds = data[dataset]
x_paper = [] #[None] * len(reviews)
x_review = [] #[None] * len(reviews)
y = [] #[None] * len(reviews)
for d in ds:
paper_content = d['paper_content']
reviews_content = d['reviews_content']
reviews = d['reviews']
for rid, (review_content,
review) in enumerate(zip(reviews_content, reviews)):
paper_ids = [
vocab[w] if w in vocab else 1
for w in paper_content.split(' ')
]
review_ids = [
vocab[w] if w in vocab else 1
for w in review_content.split(' ')
]
paper_ids = pad_sentence(paper_ids, max_len_paper, 0)
review_ids = pad_sentence(review_ids, max_len_review, 0)
xone = (paper_ids, review_ids)
yone = [np.nan] * len(aspects)
for aid, aspect in enumerate(aspects):
if aspect in review.__dict__ and review.__dict__[
aspect] is not None:
yone[aid] = float(review.__dict__[aspect])
#print rid,len(xone[0]), len(xone[1]), yone
x_paper.append(xone[0])
x_review.append(xone[1])
y.append(yone)
x_paper = np.array(x_paper, dtype=np.int32)
x_review = np.array(x_review, dtype=np.int32)
y = np.array(y, dtype=np.float32)
# add average value of missing aspect value
if fill_missing:
col_mean = np.nanmean(y, axis=0)
inds = np.where(np.isnan(y))
y[inds] = np.take(col_mean, inds[1])
print(
'Total %s dataset: %d/%d' % (dataset, len(x_paper), len(x_review)),
x_paper.shape, x_review.shape, y.shape)
data_padded.append((x_paper, x_review))
data_padded.append(y)
return data_padded, vocab, vocab_inv, label_scale, aspects
def main():
# Loading annotaions
annots = get_annots_dic()
print 'Loaded annots: %d papers and %d reviews' % (
len(annots), sum([len(v) for k, v in annots.items()]))
# Loading reviews, merging them with annotations, and saving into new directory
data_dir = "../../data/iclr_2017" # args[1] #train/reviews
datasets = ['train', 'dev', 'test']
print 'Reading reviews from...'
for dataset in datasets:
cnt_p, cnt_r = 0, 0
review_dir = os.path.join(data_dir, dataset, 'reviews_raw/')
review_annotated_dir = os.path.join(data_dir, dataset, 'reviews/')
scienceparse_dir = os.path.join(data_dir, dataset, 'scienceparse/')
model_dir = os.path.join(data_dir, dataset, 'model')
if not os.path.exists(model_dir): os.makedirs(model_dir)
review_files = sorted(glob.glob('{}/*.json'.format(review_dir)))
pids = []
for paper_json_filename in review_files:
paper = Paper.from_json(paper_json_filename)
reviews_combined = []
reviews_annotated = annots[paper.ID]
reviews_original = paper.REVIEWS
# overwrite review_annot to reviews in original paper
reviews_combined += reviews_original
for r_original in reviews_original:
r_combined = None #r_original
for r_annotated in reviews_annotated:
if r_annotated['OTHER_KEYS'] == r_original.OTHER_KEYS:
r_combined = r_original
for k, v in r_annotated.items():
if k in [
'RECOMMENDATION_UNOFFICIAL', 'SUBSTANCE',
'APPROPRIATENESS', 'MEANINGFUL_COMPARISON',
'SOUNDNESS_CORRECTNESS', 'ORIGINALITY',
'CLARITY', 'IMPACT'
]:
setattr(r_combined, k, v)
setattr(r_combined, 'IS_ANNOTATED', True)
if r_combined is None:
reviews_combined.append(r_original)
else:
reviews_combined.append(r_combined)
paper.REVIEWS = reviews_combined
cnt_r += len(paper.REVIEWS)
# save to /reviews_annotated
json.dump(
paper.to_json_object(),
open(review_annotated_dir + '/%s.json' % (paper.ID), 'w'))
print paper.ID, len(paper.REVIEWS)
cnt_p += 1
print dataset, cnt_p, cnt_r
def main(args, lower=True, max_vocab_size=False, encoder='bowtfidf'):
argc = len(args)
if argc < 9:
print(
"Usage:", args[0],
"<paper-json-dir> <scienceparse-dir> <out-dir> <submission-year> <feature output file> <tfidf vector file> <max_vocab_size> <encoder> <hand-feature>"
)
return -1
paper_json_dir = args[1] #train/reviews
scienceparse_dir = args[2] #train/parsed_pdfs
out_dir = args[3] #train/dataset
feature_output_file = args[4] #train/dataset/features.dat
vect_file = args[5] #train/dataset/vect.pkl
max_vocab_size = False if args[6] == 'False' else int(
args[6]) # False or integer
encoder = False if args[7] == 'False' else str(args[7])
hand = False if args[8] == 'False' else str(args[8])
if not os.path.exists(out_dir):
os.makedirs(out_dir)
is_train = True
vect = None
idToFeature = None
if os.path.isfile(feature_output_file):
is_train = False
idToFeature = read_features(feature_output_file)
if encoder:
print 'Loading vector file from...', vect_file
vect = load_vect(vect_file)
else:
print 'Loading vector file from scratch..'
idToFeature = dict()
outLabelsFile = open(
out_dir + '/labels_%s_%s_%s.tsv' %
(str(max_vocab_size), str(encoder), str(hand)), 'w')
outIDFile = open(
out_dir + '/ids_%s_%s_%s.tsv' %
(str(max_vocab_size), str(encoder), str(hand)), 'w')
outSvmLiteFile = open(
out_dir + '/features.svmlite_%s_%s_%s.txt' %
(str(max_vocab_size), str(encoder), str(hand)), 'w')
################################
# read reviews
################################
print 'Reading reviews from...', paper_json_dir
paper_content_corpus = [] #""
paper_json_filenames = sorted(glob.glob(
'{}/*.json'.format(paper_json_dir)))
papers = []
for paper_json_filename in paper_json_filenames:
paper = Paper.from_json(paper_json_filename)
paper.SCIENCEPARSE = ScienceParseReader.read_science_parse(
paper.ID, paper.TITLE, paper.ABSTRACT, scienceparse_dir)
paper_content_corpus.append(paper.SCIENCEPARSE.get_paper_content())
papers.append(paper)
random.shuffle(papers)
print 'Total number of reviews', len(papers)
def get_feature_id(feature):
if feature in idToFeature:
return idToFeature[feature]
else:
return None
def addFeatureToDict(fname):
id = len(idToFeature)
idToFeature[fname] = id
################################
# Initialize vocabularty
################################
outCorpusFilename = out_dir + '/corpus.pkl'
if not os.path.isfile(outCorpusFilename):
paper_content_corpus = [
preprocess(p, only_char=True, lower=True, stop_remove=True)
for p in paper_content_corpus
]
paper_content_corpus_words = []
for p in paper_content_corpus:
paper_content_corpus_words += p.split(' ')
pkl.dump(paper_content_corpus_words, open(outCorpusFilename, 'wb'))
else:
paper_content_corpus_words = pkl.load(open(outCorpusFilename, 'rb'))
print 'Total words in corpus', len(paper_content_corpus_words)
################################
# Encoding
################################
print 'Encoding..', encoder
# 1) tf-idf features on title/author_names/domains
if not encoder:
print 'No encoder', encoder
elif encoder in ['bow', 'bowtfidf']:
word_counter = Counter(paper_content_corpus_words)
# vocab limit by frequency
if max_vocab_size:
word_counter = dict(word_counter.most_common()[:max_vocab_size])
vocabulary = dict()
for w in word_counter:
if len(w) and w not in vocabulary:
if is_train:
vocabulary[w] = len(vocabulary)
addFeatureToDict(w)
else:
fid = get_feature_id(w)
if fid is not None:
vocabulary[w] = fid
print("Got vocab of size", len(vocabulary))
if is_train:
print 'Saving vectorized', vect_file
if encoder == 'bow':
vect = CountVectorizer(max_df=0.5,
analyzer='word',
stop_words='english',
vocabulary=vocabulary)
else:
vect = TfidfVectorizer(sublinear_tf=True,
max_df=0.5,
analyzer='word',
stop_words='english',
vocabulary=vocabulary)
vect.fit([p for p in paper_content_corpus])
save_vect(vect, vect_file)
# 2) sentence encoder features
elif encoder in ['w2v', 'w2vtfidf']:
from sent2vec import MeanEmbeddingVectorizer, TFIDFEmbeddingVectorizer, import_embeddings
if is_train:
w2v = import_embeddings()
vect = MeanEmbeddingVectorizer(
w2v) if encoder == 'w2v' else TFIDFEmbeddingVectorizer(w2v)
for f in range(vect.dim):
#fid = get_feature_id()
addFeatureToDict('%s%d' % (encoder, f))
print 'Saving vectorized', vect_file
if encoder == 'w2vtfidf':
vect.fit([p for p in paper_content_corpus])
save_vect(vect, vect_file)
else:
print 'Wrong type of encoder', encoder
sys.exit(1)
################################
# Add features
################################
if encoder:
all_titles = []
for p in papers:
sp = p.get_scienceparse()
title = p.get_title()
all_title = preprocess(title,
only_char=True,
lower=True,
stop_remove=True)
all_titles.append(all_title)
all_titles_features = vect.transform(all_titles)
if is_train:
print 'saving features to file', feature_output_file
if hand:
addFeatureToDict("get_most_recent_reference_year")
addFeatureToDict("get_num_references")
addFeatureToDict("get_num_refmentions")
addFeatureToDict("get_avg_length_reference_mention_contexts")
addFeatureToDict("abstract_contains_deep")
addFeatureToDict("abstract_contains_neural")
addFeatureToDict("abstract_contains_embedding")
addFeatureToDict("abstract_contains_outperform")
addFeatureToDict("abstract_contains_novel")
addFeatureToDict("abstract_contains_state_of_the_art")
addFeatureToDict("abstract_contains_state-of-the-art")
addFeatureToDict("get_num_recent_references")
addFeatureToDict("get_num_ref_to_figures")
addFeatureToDict("get_num_ref_to_tables")
addFeatureToDict("get_num_ref_to_sections")
addFeatureToDict("get_num_uniq_words")
addFeatureToDict("get_num_sections")
addFeatureToDict("get_avg_sentence_length")
addFeatureToDict("get_contains_appendix")
addFeatureToDict("proportion_of_frequent_words")
addFeatureToDict("get_title_length")
addFeatureToDict("get_num_authors")
addFeatureToDict("get_num_ref_to_equations")
addFeatureToDict("get_num_ref_to_theorems")
save_features_to_file(idToFeature, feature_output_file)
id = 1
hfws, most_frequent_words, least_frequent_words = count_words(
paper_content_corpus_words, 0.01, 0.05, 3)
for p in papers:
outIDFile.write(str(id) + "\t" + str(p.get_title()) + "\n")
rs = [r.get_recommendation() for r in p.get_reviews()]
rec = int(p.get_accepted() == True)
outLabelsFile.write(str(rec))
outSvmLiteFile.write(str(rec) + " ")
sp = p.get_scienceparse()
if encoder:
title_tfidf = all_titles_features[id - 1]
if encoder.startswith('bow'):
nz = title_tfidf.nonzero()[1]
for word_id in sorted(nz):
outSvmLiteFile.write(
str(word_id) + ":" + str(title_tfidf[0, word_id]) +
" ")
elif encoder.startswith('w2v'):
for word_id in range(vect.dim):
outSvmLiteFile.write(
str(word_id) + ":" + str(title_tfidf[word_id]) + " ")
else:
print 'wrong ecndoer', encoder
sys.exit(1)
if hand:
outSvmLiteFile.write(
str(get_feature_id("get_most_recent_reference_year")) + ":" +
str(sp.get_most_recent_reference_year() - 2000) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_num_references")) + ":" +
str(sp.get_num_references()) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_num_refmentions")) + ":" +
str(sp.get_num_refmentions()) + " ")
outSvmLiteFile.write(
str(get_feature_id(
"get_avg_length_reference_mention_contexts")) + ":" +
str(sp.get_avg_length_reference_mention_contexts()) + " ")
outSvmLiteFile.write(
str(get_feature_id("abstract_contains_deep")) + ":" +
str(int(p.abstract_contains_a_term("deep"))) + " ")
outSvmLiteFile.write(
str(get_feature_id("abstract_contains_neural")) + ":" +
str(int(p.abstract_contains_a_term("neural"))) + " ")
outSvmLiteFile.write(
str(get_feature_id("abstract_contains_embedding")) + ":" +
str(int(p.abstract_contains_a_term("embedding"))) + " ")
outSvmLiteFile.write(
str(get_feature_id("abstract_contains_outperform")) + ":" +
str(int(p.abstract_contains_a_term("outperform"))) + " ")
outSvmLiteFile.write(
str(get_feature_id("abstract_contains_novel")) + ":" +
str(int(p.abstract_contains_a_term("novel"))) + " ")
outSvmLiteFile.write(
str(get_feature_id("abstract_contains_state_of_the_art")) +
":" +
str(int(p.abstract_contains_a_term("state of the art"))) + " ")
outSvmLiteFile.write(
str(get_feature_id("abstract_contains_state-of-the-art")) +
":" +
str(int(p.abstract_contains_a_term("state-of-the-art"))) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_num_recent_references")) + ":" +
str(sp.get_num_recent_references(2017)) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_num_ref_to_figures")) + ":" +
str(sp.get_num_ref_to_figures()) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_num_ref_to_tables")) + ":" +
str(sp.get_num_ref_to_tables()) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_num_ref_to_sections")) + ":" +
str(sp.get_num_ref_to_sections()) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_num_uniq_words")) + ":" +
str(sp.get_num_uniq_words()) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_num_sections")) + ":" +
str(sp.get_num_sections()) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_avg_sentence_length")) + ":" +
str(sp.get_avg_sentence_length()) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_contains_appendix")) + ":" +
str(sp.get_contains_appendix()) + " ")
outSvmLiteFile.write(
str(get_feature_id("proportion_of_frequent_words")) + ":" +
str(
round(
sp.get_frequent_words_proportion(
hfws, most_frequent_words, least_frequent_words),
3)) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_title_length")) + ":" +
str(p.get_title_len()) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_num_authors")) + ":" +
str(sp.get_num_authors()) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_num_ref_to_equations")) + ":" +
str(sp.get_num_ref_to_equations()) + " ")
outSvmLiteFile.write(
str(get_feature_id("get_num_ref_to_theorems")) + ":" +
str(sp.get_num_ref_to_theorems()) + " ")
outSvmLiteFile.write("\n")
id += 1
outLabelsFile.close()
outIDFile.close()
outSvmLiteFile.close()
print 'saved', outLabelsFile.name
print 'saved', outIDFile.name
print 'saved', outSvmLiteFile.name
def prepare_data(
data_dir,
vocab_path='vocab',
max_vocab_size = 20000,
max_len_paper=1000,
max_len_review=200):
data_type = data_dir.split('/')[-1]
vocab_path += '.' + data_type
if max_vocab_size: vocab_path += '.'+str(max_vocab_size)
vocab_path = data_dir +'/'+ vocab_path
label_scale = 5
if 'iclr' in data_dir.lower():
fill_missing = False
aspects = ['RECOMMENDATION', 'SUBSTANCE', 'APPROPRIATENESS','MEANINGFUL_COMPARISON','SOUNDNESS_CORRECTNESS','ORIGINALITY','CLARITY', 'IMPACT', 'RECOMMENDATION_ORIGINAL']
review_dir_postfix = ''
elif 'acl' in data_dir.lower():
fill_missing = True
aspects = ['RECOMMENDATION', 'SUBSTANCE', 'APPROPRIATENESS','MEANINGFUL_COMPARISON','SOUNDNESS_CORRECTNESS','ORIGINALITY','CLARITY','IMPACT', 'REVIEWER_CONFIDENCE' ]
review_dir_postfix = ''
else:
print 'wrong dataset:',data_dir
sys.exit(1)
#Loading datasets
print 'Reading datasets..'
datasets = ['train','dev','test']
paper_content_all = []
review_content_all = []
data = defaultdict(list)
for dataset in datasets:
review_dir = os.path.join(data_dir, dataset, 'reviews%s/'%(review_dir_postfix))
scienceparse_dir = os.path.join(data_dir, dataset, 'parsed_pdfs/')
model_dir = os.path.join(data_dir, dataset, 'model/')
if not os.path.exists(model_dir): os.makedirs(model_dir)
paper_json_filenames = sorted(glob.glob('{}/*.json'.format(review_dir)))
# add all paper/review content to generate corpus for buildinb vocab
paper_content = []
review_content = []
for paper_json_filename in paper_json_filenames:
d = {}
paper = Paper.from_json(paper_json_filename)
paper.SCIENCEPARSE = ScienceParseReader.read_science_parse(paper.ID, paper.TITLE, paper.ABSTRACT, scienceparse_dir)
review_contents = []
reviews = []
for review in paper.REVIEWS:
review_contents.append(review.COMMENTS)
#preprocess(review.COMMENTS, only_char=False, lower=True, stop_remove=False))
reviews.append(review)
d['paper'] = paper
d['paper_content'] = paper.SCIENCEPARSE.get_paper_content() #preprocess( #, only_char=False, lower=True,stop_remove=False)#
d['reviews_content'] = review_contents
d['reviews'] = reviews
data[dataset].append(d)
print 'Total number of papers %d' %(np.sum([len(d) for _,d in data.items()]))
print 'Total number of reviews %d' %(np.sum([len(r['reviews']) for _,d in data.items() for r in d ]))
# Loading DATA
print 'Reading reviews from...'
data_padded = []
for dataset in datasets:
ds = data[dataset]
papers = []
x_paper = [] #[None] * len(reviews)
x_review = [] #[None] * len(reviews)
y = [] #[None] * len(reviews)
num_reviews = []
x_reviews = []
decision = []
for d in ds:
paper = d['paper']
paper_content = d['paper_content']
reviews_content = d['reviews_content']
reviews = d['reviews']
decision.append(paper.ACCEPTED)
papers.append(paper)
paper_sent = nltk.sent_tokenize(paper_content)
reviews_sent = nltk.sent_tokenize(' '.join(reviews_content))
x_paper.append(paper_sent)
x_reviews.append(reviews_sent)
num_reviews.append(len(reviews))
for rid, (review_content, review) in enumerate(zip(reviews_content,reviews)):
yone = [np.nan] * len(aspects)
review_sent = nltk.sent_tokenize(review.__dict__['COMMENTS'])
for aid,aspect in enumerate(aspects):
if aspect in review.__dict__ and review.__dict__[aspect] is not None:
yone[aid] = float(review.__dict__[aspect])
x_review.append(review_sent)
y.append(yone)
y = np.array(y, dtype=np.float32)
# add average value of missing aspect value
if fill_missing:
col_mean = np.nanmean(y,axis=0)
inds = np.where(np.isnan(y))
y[inds] = np.take(col_mean, inds[1])
data_padded.append((x_paper,papers, x_review, num_reviews, x_reviews, decision))
data_padded.append(y)
return data_padded,label_scale,aspects