def make_apsaltr_summaries(
apsal_sal_cutoff, apsal_sim_cutoff, model_summaries, max_samples=1000):
data_dir = os.path.join(os.getenv("TREC_DATA", "."), "rouge")
config_path = os.path.join(
data_dir, "apsaltr_sal_{}_sim_{}_config".format(
apsal_sal_cutoff, apsal_sim_cutoff))
config_paths = []
for job in jobs.event_cross_validation_jobs("crossval"):
for event, corpus in job.dev_events():
model_summary = model_summaries[event.fs_name()]
model_path = os.path.join(
data_dir, "model_{}".format(event.fs_name()))
max_len = len(model_summary)
df = APSalTRankSalThreshFilteredSummary().get_dataframe(
event, job.key, job.feature_set,
apsal_sal_cutoff, apsal_sim_cutoff)
updates = [update.decode('utf-8')
for update in df['text'].tolist()]
for n_sample in xrange(max_samples):
#summary_text = random_summary(updates, max_len)
sum_path = os.path.join(
data_dir, "apsaltr_sal_{}_sim_{}_sample_{}_{}".format(
apsal_sal_cutoff, apsal_sim_cutoff, n_sample,
event.fs_name()))
#with open(sum_path, 'w') as f:
# f.write(summary_text.encode('utf-8'))
config_paths.append('{} {}'.format(sum_path, model_path))
with open(config_path, 'w') as f:
f.write('\n'.join(config_paths))
return config_path
def make_ap_summaries(ap_sim_cutoff, model_summaries, max_samples=1000):
data_dir = os.path.join(os.getenv("TREC_DATA", "."), "rouge")
config_path = os.path.join(data_dir,
"ap_sim_{}_config".format(ap_sim_cutoff))
config_paths = []
for job in jobs.event_cross_validation_jobs("crossval"):
for event, corpus in job.dev_events():
model_summary = model_summaries[event.fs_name()]
model_path = os.path.join(data_dir,
"model_{}".format(event.fs_name()))
max_len = len(model_summary)
df = APFilteredSummary().get_dataframe(event, ap_sim_cutoff)
updates = [
update.decode('utf-8') for update in df['text'].tolist()
]
for n_sample in xrange(max_samples):
#summary_text = random_summary(updates, max_len)
sum_path = os.path.join(
data_dir,
"ap_sim_{}_sample_{}_{}".format(ap_sim_cutoff, n_sample,
event.fs_name()))
#with open(sum_path, 'w') as f:
# f.write(summary_text.encode('utf-8'))
config_paths.append('{} {}'.format(sum_path, model_path))
with open(config_path, 'w') as f:
f.write('\n'.join(config_paths))
return config_path
def make_apsaltr_summaries(apsal_sal_cutoff,
apsal_sim_cutoff,
model_summaries,
max_samples=1000):
data_dir = os.path.join(os.getenv("TREC_DATA", "."), "rouge")
rouge_dir = os.path.join(os.getenv("TREC_DATA", "."), "rouge", "apsaltr")
config_path = os.path.join(
data_dir,
"apsaltr_sal_{}_sim_{}_config".format(apsal_sal_cutoff,
apsal_sim_cutoff))
config_paths = defaultdict(list)
for job in jobs.event_cross_validation_jobs("crossval"):
for event, corpus in job.eval_events():
model_summary = model_summaries[event.fs_name()]
model_path = os.path.join(rouge_dir,
"model_{}".format(event.fs_name()))
max_len = len(model_summary)
event_hours = event.list_event_hours()
df = APSalTRankSalThreshFilteredSummary().get_dataframe(
event, job.key, job.feature_set, apsal_sal_cutoff,
apsal_sim_cutoff)
n_hours = len(event_hours)
for t, h in enumerate(xrange(12, n_hours, 12), 1):
print "\t", t, h
timestamp = int((event_hours[h] - \
datetime(1970,1,1)).total_seconds())
df_t = df[df['timestamp'] < timestamp]
updates = [
update.decode('utf-8') for update in df_t['text'].tolist()
]
for n_sample in xrange(max_samples):
summary_text = random_summary(updates, max_len)
sum_path = os.path.join(
data_dir,
"apsaltr_sal_{}_sim_{}_sample_{}_t{}_{}".format(
apsal_sal_cutoff, apsal_sim_cutoff, n_sample, t,
event.fs_name()))
with open(sum_path, 'w') as f:
f.write(summary_text.encode('utf-8'))
config_paths[t].append('{} {}'.format(
sum_path, model_path))
all_config_paths = []
for t in sorted(config_paths.keys()):
config_path_t = config_path + "_t{}".format(t)
print config_path_t
with open(config_path_t, 'w') as f:
f.write('\n'.join(config_paths[t]))
all_config_paths.append(config_path_t)
return all_config_paths
def make_apsaltr_summaries(
apsal_sal_cutoff, apsal_sim_cutoff, model_summaries, max_samples=1000):
data_dir = os.path.join(os.getenv("TREC_DATA", "."), "rouge")
rouge_dir = os.path.join(os.getenv("TREC_DATA", "."), "rouge", "apsaltr")
config_path = os.path.join(
data_dir, "apsaltr_sal_{}_sim_{}_config".format(
apsal_sal_cutoff, apsal_sim_cutoff))
config_paths = defaultdict(list)
for job in jobs.event_cross_validation_jobs("crossval"):
for event, corpus in job.eval_events():
model_summary = model_summaries[event.fs_name()]
model_path = os.path.join(
rouge_dir, "model_{}".format(event.fs_name()))
max_len = len(model_summary)
event_hours = event.list_event_hours()
df = APSalTRankSalThreshFilteredSummary().get_dataframe(
event, job.key, job.feature_set,
apsal_sal_cutoff, apsal_sim_cutoff)
n_hours = len(event_hours)
for t, h in enumerate(xrange(12, n_hours, 12), 1):
print "\t",t, h
timestamp = int((event_hours[h] - \
datetime(1970,1,1)).total_seconds())
df_t = df[df['timestamp'] < timestamp]
updates = [update.decode('utf-8')
for update in df_t['text'].tolist()]
for n_sample in xrange(max_samples):
summary_text = random_summary(updates, max_len)
sum_path = os.path.join(
data_dir,
"apsaltr_sal_{}_sim_{}_sample_{}_t{}_{}".format(
apsal_sal_cutoff, apsal_sim_cutoff, n_sample, t,
event.fs_name()))
with open(sum_path, 'w') as f:
f.write(summary_text.encode('utf-8'))
config_paths[t].append(
'{} {}'.format(sum_path, model_path))
all_config_paths = []
for t in sorted(config_paths.keys()):
config_path_t = config_path + "_t{}".format(t)
print config_path_t
with open(config_path_t, 'w') as f:
f.write('\n'.join(config_paths[t]))
all_config_paths.append(config_path_t)
return all_config_paths
def run_summarizer_jobs(feature_ablation=False, cross_fold=False, **kwargs):
import cuttsum.pipeline.jobs as jobs
from cuttsum.pipeline.salience import SalienceModels
if feature_ablation:
for jnum, job in enumerate(
jobs.feature_ablation_jobs(u'feature-ablation')):
print job
job.start(**kwargs)
if cross_fold:
for job in jobs.event_cross_validation_jobs("crossval"):
print job
job.start(**kwargs)
def run_summarizer_jobs(feature_ablation=False, cross_fold=False, **kwargs):
import cuttsum.pipeline.jobs as jobs
from cuttsum.pipeline.salience import SalienceModels
if feature_ablation:
for jnum, job in enumerate(
jobs.feature_ablation_jobs(u'feature-ablation')):
print job
job.start(**kwargs)
if cross_fold:
for job in jobs.event_cross_validation_jobs("crossval"):
print job
job.start(**kwargs)
def make_rank_summaries(rank_sal_cutoff,
rank_sim_cutoff,
model_summaries,
max_samples=1000):
rouge_dir = os.path.join(os.getenv("TREC_DATA", "."), "rouge")
data_dir = os.path.join(os.getenv("TREC_DATA", "."), "rouge", "rank")
if not os.path.exists(data_dir):
os.makedirs(data_dir)
config_path = os.path.join(
data_dir, "rank_sal_{}_sim_{}_config".format(rank_sal_cutoff,
rank_sim_cutoff))
config_paths = []
for job in jobs.event_cross_validation_jobs("crossval"):
for event, corpus in job.eval_events():
model_summary = model_summaries[event.fs_name()]
model_path = os.path.join(rouge_dir,
"model_{}".format(event.fs_name()))
max_len = len(model_summary)
df = RankedSalienceFilteredSummary().get_dataframe(
event, rank_sal_cutoff, rank_sim_cutoff)
print event.fs_name()
if df is None:
print "?"
import sys
sys.exit()
updates = [
update.decode('utf-8') for update in df['text'].tolist()
]
for n_sample in xrange(max_samples):
summary_text = random_summary(updates, max_len)
sum_path = os.path.join(
data_dir, "rank_sal_{}_sim_{}_sample_{}_{}".format(
rank_sal_cutoff, rank_sim_cutoff, n_sample,
event.fs_name()))
with open(sum_path, 'w') as f:
f.write(summary_text.encode('utf-8'))
config_paths.append('{} {}'.format(sum_path, model_path))
with open(config_path, 'w') as f:
f.write('\n'.join(config_paths))
return config_path
def make_rank_summaries(
rank_sal_cutoff, rank_sim_cutoff, model_summaries, max_samples=1000):
rouge_dir = os.path.join(os.getenv("TREC_DATA", "."), "rouge")
data_dir = os.path.join(os.getenv("TREC_DATA", "."), "rouge", "rank")
if not os.path.exists(data_dir):
os.makedirs(data_dir)
config_path = os.path.join(
data_dir, "rank_sal_{}_sim_{}_config".format(
rank_sal_cutoff, rank_sim_cutoff))
config_paths = []
for job in jobs.event_cross_validation_jobs("crossval"):
for event, corpus in job.eval_events():
model_summary = model_summaries[event.fs_name()]
model_path = os.path.join(
rouge_dir, "model_{}".format(event.fs_name()))
max_len = len(model_summary)
df = RankedSalienceFilteredSummary().get_dataframe(
event,
rank_sal_cutoff, rank_sim_cutoff)
print event.fs_name()
if df is None:
print "?"
import sys
sys.exit()
updates = [update.decode('utf-8')
for update in df['text'].tolist()]
for n_sample in xrange(max_samples):
summary_text = random_summary(updates, max_len)
sum_path = os.path.join(
data_dir, "rank_sal_{}_sim_{}_sample_{}_{}".format(
rank_sal_cutoff, rank_sim_cutoff, n_sample,
event.fs_name()))
with open(sum_path, 'w') as f:
f.write(summary_text.encode('utf-8'))
config_paths.append('{} {}'.format(sum_path, model_path))
with open(config_path, 'w') as f:
f.write('\n'.join(config_paths))
return config_path
shell=True)
with open(rpath, 'w') as f:
f.write(o)
recall = float(re.search('X ROUGE-2 Average_R: ([^ ]+)', o).group(1))
prec = float(re.search('X ROUGE-2 Average_P: ([^ ]+)', o).group(1))
f1 = float(re.search('X ROUGE-2 Average_F: ([^ ]+)', o).group(1))
return config_path, recall, prec, f1
model_summaries = {}
nuggets = pd.concat((cj.get_2013_nuggets(), cj.get_2014_nuggets()))
data_dir = os.path.join(os.getenv("TREC_DATA", "."), "rouge")
if not os.path.exists(data_dir):
os.makedirs(data_dir)
for job in jobs.event_cross_validation_jobs("crossval"):
for event, corpus in job.eval_events():
nugget_text = '\n'.join(
nuggets[nuggets['query id'] == event.query_id]['text'].tolist())
model_summaries[event.fs_name()] = nugget_text.decode('utf-8')
model_path = os.path.join(data_dir, "model_{}".format(event.fs_name()))
with open(model_path, 'w') as f:
f.write(nugget_text)
from cuttsum.misc import ProgressBar
apsal_sal_cutoffs = [.6]
apsal_sim_cutoffs = [.65]
apsaltr_sal_cutoffs = [.6]
apsaltr_sim_cutoffs = [.7]
shell=True)
recall = float(re.search('X ROUGE-2 Average_R: ([^ ]+)', o).group(1))
prec = float(re.search('X ROUGE-2 Average_P: ([^ ]+)', o).group(1))
f1 = float(re.search('X ROUGE-2 Average_F: ([^ ]+)', o).group(1))
return config_path, recall, prec, f1
hac_cutoffs, ap_cutoffs, apsal_cutoffs, rank_cutoffs = generate_dev_sweep()
model_summaries = {}
nuggets = pd.concat((cj.get_2013_nuggets(),
cj.get_2014_nuggets()))
data_dir = os.path.join(os.getenv("TREC_DATA", "."), "rouge")
if not os.path.exists(data_dir):
os.makedirs(data_dir)
for job in jobs.event_cross_validation_jobs("crossval"):
for event, corpus in job.dev_events():
nugget_text = '\n'.join(
nuggets[nuggets['query id'] == event.query_id]['text'].tolist())
model_summaries[event.fs_name()] = nugget_text.decode('utf-8')
model_path = os.path.join(
data_dir, "model_{}".format(event.fs_name()))
with open(model_path, 'w') as f:
f.write(nugget_text)
from cuttsum.misc import ProgressBar
rank_sal_cutoffs, rank_sim_cutoffs = rank_cutoffs
apsaltr_sal_cutoffs, apsaltr_sim_cutoffs = apsal_cutoffs
apsal_sal_cutoffs, apsal_sim_cutoffs = apsal_cutoffs
ap_sim_cutoffs = ap_cutoffs
hac_dist_cutoffs, hac_sim_cutoffs = hac_cutoffs