def get_opinosis_df(test_df, dev_df, config):
opinosis_df = test_df.copy()
max_summary_list = [4, 5, 6]
redundancy_list = [0, 1]
gap_list = [2, 3, 4]
run_id_rougeL = {}
for max_summary in max_summary_list:
for redundancy in redundancy_list:
for gap in gap_list:
run_id = 'max-%i-red-%i-gap-%i' % (max_summary, redundancy,
gap)
dir_dev = os.path.join('eval', config.data, 'opinosis', 'dev',
'output', run_id)
paths_dev = [
os.path.join(dir_dev,
'{0:03}.{1}.system'.format(i, run_id))
for i in dev_df.index
]
summary_list = []
for path_dev in paths_dev:
with open(path_dev, 'r') as f:
summary = refine_text(f.read())
summary_list += [summary]
rougeL = compute_rouge(dev_df, list(dev_df.summary),
summary_list).mean()['rougeL']
run_id_rougeL[run_id] = rougeL
run_id_test = sorted(run_id_rougeL.items(),
key=lambda x: x[1],
reverse=True)[0][0]
dir_test = os.path.join('eval', config.data, 'opinosis', 'test', 'output',
run_id_test)
paths_test = [
os.path.join(dir_test, '{0:03}.{1}.system'.format(i, run_id_test))
for i in opinosis_df.index
]
summary_list = []
for path_test in paths_test:
with open(path_test, 'r') as f:
summary = refine_text(f.read())
summary_list += [summary]
opinosis_df[['rouge1', 'rouge2',
'rougeL']] = compute_rouge(opinosis_df,
list(opinosis_df.summary),
summary_list)
opinosis_df.to_pickle(config.path_opinosis)
return opinosis_df
def get_denoisesum_df(config):
denoisesum_df = pd.read_json(config.path_denoisesum_raw)
denoisesum_df[['rouge1', 'rouge2', 'rougeL']] = compute_rouge(
denoisesum_df,
reference_list=list(denoisesum_df.summary),
summary_list=list(denoisesum_df.denoisesum))
return denoisesum_df
def get_lead_df(data_df, config):
lead_df = data_df.copy()
lead_df['lead'] = lead_df.text.apply(lambda text: ' '.join(
[sent_tokenize(doc)[0].strip() for doc in text.split('</DOC>')]))
lead_df[['rouge1', 'rouge2', 'rougeL']] = \
compute_rouge(lead_df, reference_list=list(lead_df.summary), summary_list=list(lead_df.lead))
lead_df.to_pickle(config.path_lead)
return lead_df
def get_meansum_df(config):
meansum_df = pd.read_pickle(config.path_meansum_raw)
meansum_df[['rouge1', 'rouge2', 'rougeL'
]] = compute_rouge(meansum_df,
reference_list=list(meansum_df.summary),
summary_list=list(meansum_df.meansum))
meansum_df.to_pickle(config.path_meansum)
return meansum_df
def get_recursum_df(sess, model, test_df, sys_sum, topk, threshold, truncate,
summary_l, num_split):
data_df = test_df.groupby('business_id').agg({
'doc_l':
lambda doc_l_series: doc_l_series.values[0],
'sent_l':
lambda sent_l_series: sent_l_series.values[0],
'token_idxs':
lambda token_idxs_series: token_idxs_series.values[0],
'text':
lambda text_series: text_series.values[0]
})
batches = get_batches(data_df, model.config.batch_size)
topic_sents_list, probs_topic_list, topic_tokens_list = compute_topic_sents_probs(
sess, model, batches, mode='eval', sample=False)
text_list = [row.text.replace('\n', '') for _, row in data_df.iterrows()]
verbose = False
args = [(model.config.tree_idxs, topic_sents, text, topk, threshold,
truncate, summary_l, verbose)
for topic_sents, text in zip(topic_sents_list, text_list)]
pool = multiprocessing.Pool(processes=num_split)
summary_l_sents_list = pool.map(sys_sum, args)
pool.close()
summary_list = [
get_text_from_sents(summary_l_sents[summary_l]['sents'])
for summary_l_sents in summary_l_sents_list
]
summary_idxs_list = [[model.config.topic_idxs[topic_index] for topic_index in summary_l_sents[summary_l]['indices']] \
for summary_l_sents in summary_l_sents_list]
data_df['recursum'] = summary_list
data_df['summary_idxs'] = summary_idxs_list
data_df['topic_sents'] = topic_sents_list
data_df['topic_tokens'] = topic_tokens_list
data_df['probs_topic'] = probs_topic_list
recursum_df = pd.merge(test_df,
data_df[[
'recursum', 'summary_idxs', 'topic_sents',
'topic_tokens', 'probs_topic'
]],
on='business_id',
how='left')
recursum_df[['rouge1', 'rouge2', 'rougeL']] = compute_rouge(recursum_df, \
reference_list=list(recursum_df.summary), summary_list=list(recursum_df.recursum))
recursum_df = recursum_df.set_index(test_df.index)
assert recursum_df['business_id'].to_dict(
) == test_df['business_id'].to_dict()
return recursum_df
def get_copycat_df(data_df, config):
business_id_dict = {
'#NAME1': '-zbcosKSMGDhaZYN-CrcVA',
'#NAME2': '-i3pCgQi_Y9NiSSWs6G7bw',
'#NAME3': '-_TSaVr53qiEGqMkwyEMaQ',
'#NAME4': '-vCLrTTgw6pBufdarW8ynA',
'#NAME5': '-K3kqmykKlhlB4arCsLHOw',
'#NAME6': '-exEWEQ3iSMVC-QUP_ycPQ',
'#NAME7': '-_yEVC3_3M6YOsamYfNFEw',
'#NAME8': '-NR4KqS6lHseNvJ-GFzfMA',
'#NAME9': '-ot4Xd6GxSUOqwUj7okZuA',
'#NAME10': '-pV9kWNoA9vyHfM_auYecA',
'#NAME11': '-FNquqGseSCVMWo7KbK-Tg',
'#NAME12': '-Qkx7W0itbAApcG5lJuMFQ',
'#NAME13': '-SJcjOv88ZHjIU44U4vWTQ',
'#NAME14': '-isxnIljKLVjc9qEhCiaGg',
'#NAME15': '-iPc_YSSqvM1CpZxxeUTXw',
'#NAME16': '-ADtl9bLp8wNqYX1k3KuxA',
'#NAME17': '-zEpEmDfFQL-ph0N3BDlXA',
'#NAME18': '-oOKqZbYDt08zaWWyLZNIw',
'#NAME19': '-PbCfkydmvuNcG9VG_ixkQ',
'#NAME20': '-pN44P-_PjRpcj4Rk2wMOg',
'#NAME21': '-dcI8oWvxdMCGp00da8Ksg',
'#NAME22': '-MKWJZnMjSit406AUKf7Pg'
}
json_copycat = json.load(open(config.path_copycat_raw))
json_copycat_df = pd.concat([
pd.DataFrame.from_dict(json_copycat[category], orient='index')
for category in list(json_copycat)[:-1]
])
if config.data == 'yelp':
json_copycat_df.index = [
business_id_dict[raw_id] if 'NAME' in raw_id else raw_id
for raw_id in json_copycat_df.index
]
copycat_df = pd.merge(data_df, \
pd.DataFrame([{'business_id': index, 'copycat': ' '.join(copycat_summary)} \
for index, copycat_summary_list in json_copycat_df.gen_summ.to_dict().items() \
for copycat_summary in copycat_summary_list]), how='left')
copycat_df[['rouge1', 'rouge2', 'rougeL'
]] = compute_rouge(copycat_df,
reference_list=list(copycat_df.summary),
summary_list=list(copycat_df.copycat))
copycat_df.index = data_df.index
assert len(data_df) == len(copycat_df)
copycat_df.to_pickle(config.path_copycat)
return copycat_df
def get_lexrank_df(data_df, n_sents, min_sent_l, config):
def lexrank(tfidfbows):
cos_matrix = cosine_similarity(tfidfbows.toarray(),
tfidfbows.toarray())
eig_values, _ = np.linalg.eig(cos_matrix)
eig_indices = np.argsort(eig_values)[::-1]
return eig_indices
lexrank_df = data_df.copy()
lexrank_df['eig_indices'] = lexrank_df['tfidfbows'].apply(lexrank)
lexrank_df['length_indices'] = data_df.sent_l.apply(
lambda sent_l: np.where(np.array(sent_l) > min_sent_l)[0])
lexrank_df['summary_indices'] = lexrank_df.apply(
lambda row: [i for i in row.eig_indices
if i in row.length_indices][:n_sents], 1)
lexrank_df['lexrank'] = lexrank_df.apply(lambda row: \
get_text_from_sents(idxs_to_sents(np.array(row.token_idxs)[row.summary_indices], config)), 1)
lexrank_df[['rouge1', 'rouge2', 'rougeL']] = compute_rouge(lexrank_df, \
reference_list=list(lexrank_df.summary), summary_list=list(lexrank_df.lexrank))
lexrank_df.to_pickle(config.path_lexrank)
return lexrank_df