def load_datasets_tira_evaluation(test_dataset_main_directory, preset_key):
"""Load the PAN dataset for **Tira** evaluation.
This function loads the PAN training and test dataset and truth by calling the *ProcessDataFiles* module twice,
then passes them along with Author IDs of the test dataset.
"""
# Define the dictionary of presets. Each “preset” is a dictionary of some values.
PRESETS_DICTIONARY = {
'PAN18_English': {
'dataset_name': 'PAN 2018 English',
'xmls_subdirectory': '/en/',
'truth_subpath': '/en-truth/truth.txt',
}
}
PRESET = PRESETS_DICTIONARY[preset_key]
# Define the constant and the paths
TRAINING_DATASET_MAIN_DIRECTORY =\
"U:/TA"
# # TEMP (TIRA): For local testing on TIRA
# TRAINING_DATASET_MAIN_DIRECTORY = "E:/author-profiling/pan18-author-profiling-training-dataset-2018-02-27"
xmls_directory_train = os.path.join(TRAINING_DATASET_MAIN_DIRECTORY,
PRESET['xmls_subdirectory'])
truth_path_train = os.path.join(TRAINING_DATASET_MAIN_DIRECTORY,
PRESET['truth_subpath'])
xmls_directory_test = os.path.join(test_dataset_main_directory,
PRESET['xmls_subdirectory'])
# ↳ Note: truth_path_test will not be provided to the participants.
# Load the PAN 2018 training dataset and truth from the files into lists
print("Loading the %s training dataset and truth...",
PRESET['dataset_name'])
docs_train, y_train, author_ids_train, original_tweet_lengths_train = \
process_data_files.load_pan_data(xmls_directory_train, truth_path_train, False, None)
# Load the PAN 2018 test dataset from the files into lists
print("Loading the %s test dataset...", PRESET['dataset_name'])
docs_test, y_test, author_ids_test, original_tweet_lengths_test = \
process_data_files.load_pan_data(xmls_directory_test, None, False, None)
# ↳ Note: truth_path_test will not be provided to the participants. As a result, *truths_test* will be empty.
return docs_train, docs_test, y_train, author_ids_test
def load_datasets_development(preset_key):
"""Load the PAN dataset for the development phase.
This function loads the PAN training dataset and truth by calling the *ProcessDataFiles* module,
then splits the dataset into training and test sets.
"""
directory = 'U:/TA'
# Define the dictionary of presets. Each “preset” is a dictionary of some values.
PRESETS_DICTIONARY = {
'PAN18_English': {
'dataset_name': 'PAN 2018 English',
'xmls_directory': directory + '/en/',
'truth_path': directory + '/truth/truth.txt',
'txts_destination_directory': directory,
}
}
PRESET = PRESETS_DICTIONARY[preset_key]
# Load the PAN 2018 training dataset and the truth from the files into lists
print("Loading the %s training dataset and the truth...",
PRESET['dataset_name'])
merged_tweets_of_authors, truths, author_ids, original_tweet_lengths =\
process_data_files.load_pan_data(PRESET['xmls_directory'], PRESET['truth_path'],
False, PRESET['txts_destination_directory'])
# Split the dataset into balanced (stratified) training and test sets:
docs_train, docs_test, y_train, y_test, author_ids_train, author_ids_test,\
original_tweet_lengths_train, original_tweet_lengths_test =\
train_test_split(merged_tweets_of_authors, truths, author_ids, original_tweet_lengths,
test_size=0.4, random_state=42, stratify=truths)
# ↳ *stratify=truths* selects a balanced sample from the data, with the same class proportion as the *truths* list.
# • Sort all lists in the ascending order of *author_ids* (separately, for the training and test set)
# This is only done for the sakes of consistency between the *load_datasets_development()* and
# *load_datasets_tira_evaluation()* functions, because the output of the latter is sorted by *author_ids*, while the
# former is shuffled by the *train_test_split()* function.
# Sort the training set
author_ids_train, docs_train, y_train, original_tweet_lengths_train = [
list(tuple) for tuple in zip(*sorted(
zip(author_ids_train, docs_train, y_train,
original_tweet_lengths_train)))
]
# Sort the test set
author_ids_test, docs_test, y_test, original_tweet_lengths_test = [
list(tuple) for tuple in zip(*sorted(
zip(author_ids_test, docs_test, y_test,
original_tweet_lengths_test)))
]
# # TEMP: Used for producing a mimic of the **TIRA** environment
# ProcessDataFiles.split_train_and_test_files(author_ids_train, author_ids_test, y_train, y_test, preset_key)
return docs_train, docs_test, y_train, y_test
