def main():
runs = ['3', '4']
#sources = ['raw_processed', 'clean_contradictions_batch_0.5']
sources = [
'data_processed', 'clean_contradictions_total_1',
'clean_contradictions_pair_0.5', 'clean_contradictions_batch_0.5'
]
group = 'experiment*'
n_q = '*'
n_lists = '*'
batch = '*'
#name = f'run{"_".join(runs)}-group_{group}-batch{batch}'.replace('*', '-all-')
n_lists = '*'
for source in sources:
data_dict_list = []
name = f'run{"_".join(runs)}-group_{group}-batch{batch}'.replace(
'*', '-all-')
name = f'{name}-{source}'
print(name)
for run in runs:
print(run)
data = load_processed_data(run, group, n_q, n_lists, batch, source)
data_dict_list.extend(data)
# data_dict_list.extend(load_experiment_data(run, group, n_q, n_lists, batch, remove_not_val = True))
print(len(data_dict_list))
print('checking if concepts are there:')
check_data(data_dict_list)
print('creating input')
input_df = create_input_df(data_dict_list)
print(input_df.columns)
input_dir = '../analyses/crowdtruth/input/'
input_path = f'{input_dir}{name}.csv'
os.makedirs(input_dir, exist_ok=True)
input_df.to_csv(input_path, index=False)
res_dir = '../analyses/crowdtruth/results/'
res_path = f'{res_dir}{name}'
os.makedirs(res_dir, exist_ok=True)
print('running crowdtruth')
input_file = input_path
data, config = crowdtruth.load(file=input_file, config=TestConfig())
results = crowdtruth.run(data, config)
print('crowdtruth done')
unit_scores = results['units']
split_unit_annotation_score(unit_scores)
unit_scores.to_csv(f'{res_path}-units.csv')
worker_scores = results['workers']
worker_scores.to_csv(f'{res_path}-workers.csv')
annotation_scores = results["annotations"]
annotation_scores.to_csv(f'{res_path}-annotations.csv')
print(f'results stored: {res_path}')
def test_all_workers_agree(self):
for w in range(2,11):
test_config = self.test_conf_const.__class__
data, config = crowdtruth.load(
file=TEST_FILE_PREF + str(w) + "work_agr.csv",
config=test_config())
results = crowdtruth.run(data, config)
self.assertAlmostEqual(results["units"]["uqs"].at[1], 1.0)
for wid in range(w):
self.assertAlmostEqual(results["workers"]["wqs"].at["W" + str(wid + 1)], 1.0)
if not config.open_ended_task:
self.assertAlmostEqual(results["annotations"]["aqs"]["A"], 1.0)
def test_metrics_correct_interval(self):
test_conf_const = TutorialCustomizedConfig()
test_config = test_conf_const.__class__
data, config = crowdtruth.load(
file="tutorial/relex_example_custom.csv",
config=test_config())
results = crowdtruth.run(data, config)
for _, val_arr in results["units"]["unit_annotation_score"].items():
for _, val in val_arr.items():
self.assertGreaterEqual(val, 0.0)
self.assertLessEqual(val, 1.0)
for _, val in results["units"]["uqs"].items():
self.assertGreaterEqual(val, 0.0)
self.assertLessEqual(val, 1.0)
for _, val in results["workers"]["wqs"].items():
self.assertGreaterEqual(val, 0.0)
self.assertLessEqual(val, 1.0)
for _, val in results["annotations"]["aqs"].items():
self.assertGreaterEqual(val, 0.0)
self.assertLessEqual(val, 1.0)
def create_analysis_files(dataset, max_no_workers, max_runs, storing_folder):
unique_unit_ids = get_uniq_unit_ids(dataset)
for subset_size in xrange(3, max_no_workers):
workers_directory = storing_folder + str(subset_size) + "workers"
if not os.path.exists(workers_directory):
os.makedirs(workers_directory)
map_unit_id_combinations = {}
for unit_id in xrange(0, len(unique_unit_ids)):
(count,
subset_unit_id) = get_no_work_unit_id(dataset,
unique_unit_ids[unit_id])
combinations = gen_all_worker_combinations(subset_size, count,
subset_unit_id)
map_unit_id_combinations[unique_unit_ids[unit_id]] = combinations
import csv
fields = [
'F1_nr_binary', 'F1_n_binary', 'F1_nr_ternary', 'F1_r_ternary',
'F1_hr_ternary'
]
with open('F1_' + str(subset_size) + ".csv", 'w') as f:
writer = csv.writer(f)
writer.writerow(fields)
for run_no in xrange(1, max_runs):
unit_worker_set = {}
for unit_id, worker_sets in map_unit_id_combinations.iteritems():
unit_worker_set[unit_id] = pick_random_worker_set(worker_sets)
df_subset_size = pd.DataFrame()
for unit_id, worker_set in unit_worker_set.iteritems():
df_subset = dataset[(dataset["_unit_id"] == unit_id)
& (dataset["_worker_id"].isin(worker_set))]
frames = [df_subset_size, df_subset]
df_subset_size = pd.concat(frames)
df_subset_size.to_csv(workers_directory + "/run.csv", index=False)
filename = workers_directory + "/run.csv"
results_with_newGT = pd.read_csv(
"../ground_truth_data/reviewers_pilot_aggregated_judgments.csv"
)
class config(DefaultConfig):
inputColumns = [
"index", "bin", "doc_len", "document_id", "document_body",
"document_title", "rel", "topic", "topic_description",
"topic_query"
]
outputColumns = ["relevant_snippets"]
# processing of a closed task
open_ended_task = True
annotation_separator = ","
annotation_vector = []
def processJudgments(self, judgments):
for col in self.outputColumns:
judgments[col] = judgments[col].apply(
lambda x: x.replace('[', ''))
judgments[col] = judgments[col].apply(
lambda x: x.replace(']', ''))
judgments[col] = judgments[col].apply(
lambda x: x.replace('"', ''))
judgments[col] = judgments[col].apply(
lambda x: x.replace(' ', ','))
return judgments
# Read data
data, config = crowdtruth.load(file=filename, config=config())
data['judgments'].head()
results = crowdtruth.run(data, config)
results["units"]['max_relevance_score'] = pd.Series(
np.random.randn(len(results["units"])),
index=results["units"].index)
for i in xrange(0, len(results["units"])):
maxVal = 0.0
for key, value in results["units"][
"unit_annotation_score"].iloc[i].items():
if key != "none":
if value > maxVal:
maxVal = value
results["units"]['max_relevance_score'].iloc[i] = maxVal
results["units"]["reviewers_rel"] = pd.Series(
np.random.randn(len(results["units"])),
index=results["units"].index)
results["units"]["reviewers_rel_merged"] = pd.Series(
np.random.randn(len(results["units"])),
index=results["units"].index)
for i in xrange(0, len(results_with_newGT.index)):
for j in xrange(0, len(results["units"].index)):
if (results_with_newGT["topic"].iloc[i]
== results["units"]["input.topic"].iloc[j]) and (
results_with_newGT["document_id"].iloc[i] ==
results["units"]["input.document_id"].iloc[j]):
results["units"]["reviewers_rel"].iloc[
j] = results_with_newGT["reviewers_rel"].iloc[i]
results["units"]["reviewers_rel_merged"].iloc[
j] = results_with_newGT[
"reviewers_rel_merged"].iloc[i]
F1_notrelevant_binary = compute_F1_score_not_relevant_binary(
results["units"])
F1_relevant_binary = compute_F1_score_relevant_binary(
results["units"])
F1_notrelevant_ternary = compute_F1_score_not_relevant_ternary(
results["units"])
F1_relevant_ternary = compute_F1_score_relevant_ternary(
results["units"])
F1_highlyrelevant_ternary = compute_F1_score_highly_relevant_ternary(
results["units"])
row = [
F1_notrelevant_binary, F1_relevant_binary,
F1_notrelevant_ternary, F1_relevant_ternary,
F1_highlyrelevant_ternary
]
with open('F1_' + str(subset_size) + ".csv", 'a') as f:
writer = csv.writer(f)
writer.writerow(row)
def test_incremental_worker_agreement(self):
for w in range(4, 11):
test_config = self.test_conf_const.__class__
data, config = crowdtruth.load(
file=TEST_FILE_PREF + str(w - 2) + "vs" + str(w - 1) + "work_agr.csv",
config=test_config())
results = crowdtruth.run(data, config)
# print str(config.open_ended_task)
# check that workers that agree on the same unit have the same quality score
for x in range(2, w):
if x != (w - 1):
self.assertAlmostEqual(
results["workers"]["wqs"].at["W1"],
results["workers"]["wqs"].at["W" + str(x)],)
self.assertAlmostEqual(
results["workers"]["wqs"].at["W" + str(w)],
results["workers"]["wqs"].at["W" + str(w + x - 1)])
# workers that agree have a greater WQS than the worker that disagrees
self.assertGreater(
results["workers"]["wqs"].at["W1"],
results["workers"]["wqs"].at["W" + str(w - 1)])
self.assertGreater(
results["workers"]["wqs"].at["W" + str(w)],
results["workers"]["wqs"].at["W" + str(2 * w - 1)])
# the more workers agree on a unit, the higher the worker quality score
self.assertGreater(
results["workers"]["wqs"].at["W" + str(w)],
results["workers"]["wqs"].at["W1"])
# print "W" + str(w) + ": " + str(results["workers"]["wqs"].at["W" + str(w)])
# print "W1: " + str(results["workers"]["wqs"].at["W1"])
# the more workers agree on a unit, the higher the unit quality score
self.assertLess(
results["units"]["uqs"].at[1],
results["units"]["uqs"].at[2])
self.assertLess(
results["units"]["uqs"].at[1],
results["units"]["uqs"].at[3])
self.assertLess(
results["units"]["uqs"].at[2],
results["units"]["uqs"].at[3])
# the more workers agree on an annotation, the higher the unit quality score
if not config.open_ended_task:
self.assertLess(
results["annotations"]["aqs"].at["A"],
results["annotations"]["aqs"].at["C"])
self.assertLess(
results["annotations"]["aqs"].at["B"],
results["annotations"]["aqs"].at["A"])
self.assertLess(
results["annotations"]["aqs"].at["D"],
results["annotations"]["aqs"].at["C"])
self.assertLess(
results["annotations"]["aqs"].at["A"],
results["annotations"]["aqs"].at["E"])
self.assertLess(
results["annotations"]["aqs"].at["C"],
results["annotations"]["aqs"].at["E"])