def main():
# Load in the data
prompts = np.asarray(preprocessor.read_prompts())
annotations = np.asarray(preprocessor.read_annotations())
data = {}
preds, y_test = [], []
# store data in dictionary
for p in prompts:
data[p[0]] = {'xml': p[1], 'outcome': p[2], 'intervention': p[3], 'comparator': p[4], 'answer': '', 'reasoning': ''}
for a in annotations:
if (a[3]):
data[a[1]]['answer'] = a[7]
if (a[4]):
data[a[1]]['reasoning'] += str(a[6])
test_id = preprocessor.test_document_ids()
# get predictions and add them to array
for k in data.keys():
# try to parse text to remove weird things
id_ = data[k]['xml']
# if the file is not a test file
if not(id_ in test_id):
continue
out = try_except_parse(data[k]['outcome'])
inter = try_except_parse(data[k]['intervention'])
cmp = try_except_parse(data[k]['comparator'])
ans = try_except_parse(data[k]['answer'])
res = try_except_parse(data[k]['reasoning'])
if (ans == ''):
continue # we don't have a valid answer for this one...
y_test.append(ans)
# just use the reasoning as our sentence
likely_sentence = res
guess = eval_sentence(likely_sentence, out, inter, cmp)
if (guess == "No significant difference"):
preds.append(0)
elif (guess == "Significantly decreased"):
preds.append(-1)
else:
preds.append(1)
acc = accuracy_score(y_test, preds)
f1 = f1_score(y_test, preds, average='macro')
prec = precision_score(y_test, preds, average = 'macro')
rec = recall_score(y_test, preds, average = 'macro')
return acc, f1, prec, rec
def load_data(use_test, bow=True):
"""
Load the data into a train/val/test set that allows for easy access.
@return bag-of-word representation of training, validation, test sets (with labels).
"""
prompts = preprocessor.read_prompts()
annotations = preprocessor.read_annotations()
# filter out prompts for which we do not have annotations for whatever reason
# this was actually just one case; not sure what was going on there.
def have_annotations_for_prompt(prompt_id):
return len(
annotations[annotations[PROMPT_ID_COL_NAME] == prompt_id]) > 0
prompts = [
prompt for row_idx, prompt in prompts.iterrows()
if have_annotations_for_prompt(prompt[PROMPT_ID_COL_NAME])
]
prompts = pd.DataFrame(prompts)
# Sort into training and validation by article id
train_doc_ids = preprocessor.train_document_ids()
val_doc_ids = preprocessor.validation_document_ids()
test_doc_ids = preprocessor.test_document_ids()
# get a dev set randomly
dev_doc_ids = list(train_doc_ids)
random.shuffle(dev_doc_ids)
dev_doc_ids = set(dev_doc_ids[:int(len(dev_doc_ids) * .1)])
x_train, y_train, x_dev, y_dev, x_val, y_val, x_test, y_test = [], [], [], [], [], [], [], []
pids = prompts[STUDY_ID_COL].values
for i in range(len(pids)):
annotations_for_prompt = annotations[annotations[PROMPT_ID_COL_NAME] ==
prompts["PromptID"].values[i]]
labels = annotations_for_prompt[[LBL_COL_NAME,
EVIDENCE_COL_NAME]].values
id_ = pids[i]
# this is all of the reasonings
articles = [a[1] for a in labels]
for article_text in articles:
# extract i/c/o
out = prompts["Outcome"].values[i].lower()
inter = prompts["Intervention"].values[i].lower()
cmp = prompts["Comparator"].values[i].lower()
# add to correct pile: train/val/test
tmp = [article_text, out, inter, cmp]
loss = stats.mode([l1[0] for l1 in labels])[0][0]
if id_ in dev_doc_ids and not (use_test):
x_dev.append(tmp)
y_dev.append(loss)
elif id_ in train_doc_ids:
x_train.append(tmp)
y_train.append(loss)
elif id_ in val_doc_ids:
x_val.append(tmp)
y_val.append(loss)
elif id_ in test_doc_ids:
x_test.append(tmp)
y_test.append(loss)
else:
raise ValueError("Unknown study id {}".format(id_))
# transform to np.array
y_test = np.asarray(y_test)
# if we are removing the test set, use validation as test set.
if not (use_test):
x_test = x_val
y_test = y_val
x_val = x_dev
y_val = y_dev
print("Running bag of words...")
ret = bag_of_words(
x_train, y_train, x_val, y_val, x_test,
y_test) if bow else [x_train, y_train, x_val, y_val, x_test, y_test]
return ret
def load_data(use_test, bow=True):
"""
Load the data into a train/val/test set that allows for easy access.
@return bag-of-word representation of training, validation, test sets (with labels).
"""
prompts = preprocessor.read_prompts()
annotations = preprocessor.read_annotations()
# filter out prompts for which we do not have annotations for whatever reason
# this was actually just one case; not sure what was going on there.
def have_annotations_for_prompt(prompt_id):
return len(
annotations[annotations[PROMPT_ID_COL_NAME] == prompt_id]) > 0
prompts = [
prompt for row_idx, prompt in prompts.iterrows()
if have_annotations_for_prompt(prompt[PROMPT_ID_COL_NAME])
]
prompts = pd.DataFrame(prompts)
# Sort into training and validation by article id
train_doc_ids = preprocessor.train_document_ids()
val_doc_ids = preprocessor.validation_document_ids()
test_doc_ids = preprocessor.test_document_ids()
# get a dev set randomly
dev_doc_ids = list(train_doc_ids)
random.shuffle(dev_doc_ids)
dev_doc_ids = set(dev_doc_ids[:int(len(dev_doc_ids) * .1)])
x_train, y_train, x_dev, y_dev, x_val, y_val, x_test, y_test = [], [], [], [], [], [], [], []
pids = prompts[STUDY_ID_COL].values
for i in range(len(pids)):
id_, data, losses = parse_prompt_id_data(annotations, prompts, pids, i)
in_training = id_ in (train_doc_ids - dev_doc_ids)
# get a reasoning from previous/next prompt id
if (i > 0 and id_ == pids[i - 1] and not (in_training)):
_, mismatched_data, _ = parse_prompt_id_data(
annotations, prompts, pids, i - 1)
# add the mismatched data here
row = copy.deepcopy(data[0])
row[1] = mismatched_data[0][1]
data.append(row)
losses.append(losses[-1])
elif (i < len(pids) and id_ == pids[i + 1] and not (in_training)):
_, mismatched_data, _ = parse_prompt_id_data(
annotations, prompts, pids, i + 1)
# add the mismatched data here
row = copy.deepcopy(data[0])
row[1] = mismatched_data[0][1]
data.append(row)
losses.append(losses[-1])
for i in range(len(data)):
tmp = data[i]
loss = losses[i]
# find where to put this section
if id_ in dev_doc_ids and not (use_test):
x_dev.append(tmp)
y_dev.append(loss)
elif id_ in train_doc_ids:
x_train.append(tmp)
y_train.append(loss)
elif id_ in val_doc_ids:
x_val.append(tmp)
y_val.append(loss)
elif id_ in test_doc_ids:
x_test.append(tmp)
y_test.append(loss)
else:
raise ValueError("Unknown study id {}".format(id_))
# if we are removing the test set, use validation as test set.
if not (use_test):
x_test = x_val
y_test = y_val
x_val = x_dev
y_val = y_dev
ret = bag_of_words(
x_train, y_train, x_val, y_val, x_test, y_test,
5) if bow else [x_train, y_train, x_val, y_val, x_test, y_test]
return ret
min_span = ''
for i_offset in range(-search_range, search_range):
for f_offset in range(-search_range, search_range):
span = text[i + i_offset:f + f_offset]
dist = string_distance(ev.strip(' '), span.strip(' '))
if dist <= min_dist:
min_dist = dist
min_span = span
if min_span:
return True
return False
annotations = pp.read_annotations()
counter = 0
almost = 0
exact = 0
total = len(annotations)
for _, annot in annotations.iterrows():
article_file = path.join(DATA_DIR,
"txt_files/PMC" + str(annot.PMCID) + ".txt")
with open(article_file, encoding='utf-8') as f:
text = f.read()
start, end = annot["Evidence Start"], annot["Evidence End"]
raw_text = text[start:end + 1]
saved_text = pp.extract_raw_text(pp.get_article(annot.PMCID))[start:end +