import os
import time
import random
from enum import Enum
from dataclasses import dataclass, field
from typing import List, Optional, Union, Dict, Tuple
import torch
from torch.utils.data.dataset import Dataset
from filelock import FileLock
import logging
from transformers import (InputFeatures,
InputExample,
PreTrainedTokenizerBase)
import pandas as pd
from sklearn.metrics import precision_recall_fscore_support
import sys
sys.path.append("../")
sys.path.append('./biobert_re/')
from data_processor import glue_convert_examples_to_features, glue_output_modes, glue_processors
import utils
from ehr import HealthRecord
from annotations import Relation
logger = logging.getLogger(__name__)
@dataclass
class GlueDataTrainingArguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
Using `HfArgumentParser` we can turn this class into argparse arguments to be able to specify them on the command
line.
"""
task_name: str = field(metadata={"help": "The name of the task to train on: " + ", ".join(glue_processors.keys())})
data_dir: str = field(
metadata={"help": "The input data dir. Should contain the .tsv files (or other data files) for the task."}
)
max_seq_length: int = field(
default=128,
metadata={
"help": "The maximum total input sequence length after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded."
},
)
overwrite_cache: bool = field(
default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
)
def __post_init__(self):
self.task_name = self.task_name.lower()
class Split(Enum):
train = "train"
dev = "dev"
test = "test"
# noinspection PyTypeChecker
class REDataset(Dataset):
"""
A class representing a training dataset for Relation Extraction.
"""
args: GlueDataTrainingArguments
output_mode: str
features: List[InputFeatures]
def __init__(
self,
args: GlueDataTrainingArguments,
tokenizer: PreTrainedTokenizerBase,
limit_length: Optional[int] = None,
mode: Union[str, Split] = Split.train,
cache_dir: Optional[str] = None,
):
self.args = args
self.processor = glue_processors[args.task_name]()
self.output_mode = glue_output_modes[args.task_name]
if isinstance(mode, str):
try:
mode = Split[mode]
except KeyError:
raise KeyError("mode is not a valid split name")
# Load data features from cache or dataset file
cached_features_file = os.path.join(
cache_dir if cache_dir is not None else args.data_dir,
"cached_{}_{}_{}_{}".format(
mode.value,
tokenizer.__class__.__name__,
str(args.max_seq_length),
args.task_name,
),
)
label_list = self.processor.get_labels()
self.label_list = label_list
# Make sure only the first process in distributed training processes the dataset,
# and the others will use the cache.
lock_path = cached_features_file + ".lock"
with FileLock(lock_path):
if os.path.exists(cached_features_file) and not args.overwrite_cache:
start = time.time()
self.features = torch.load(cached_features_file)
logger.info(f"Loading features from cached file {cached_features_file} [took %.3f s]", time.time() - start)
else:
logger.info(f"Creating features from dataset file at {args.data_dir}")
if mode == Split.dev:
examples = self.processor.get_dev_examples(args.data_dir)
elif mode == Split.test:
examples = self.processor.get_test_examples(args.data_dir)
else:
examples = self.processor.get_train_examples(args.data_dir)
if limit_length is not None:
examples = examples[:limit_length]
self.features = glue_convert_examples_to_features(
examples,
tokenizer,
max_length=args.max_seq_length,
label_list=label_list,
output_mode=self.output_mode,
)
start = time.time()
torch.save(self.features, cached_features_file)
logger.info("Saving features into cached file %s [took %.3f s]", cached_features_file, time.time() - start)
def __len__(self):
return len(self.features)
def __getitem__(self, i) -> InputFeatures:
return self.features[i]
def get_labels(self):
return self.label_list
class RETestDataset(Dataset):
"""
A class representing a test Dataset for relation extraction.
"""
def __init__(self, test_ehr, tokenizer, max_seq_len, label_list):
self.re_text_list, self.relation_list = generate_re_test_file(test_ehr)
if not self.re_text_list:
self.features = []
else:
examples = []
for (i, text) in enumerate(self.re_text_list):
guid = "%s" % i
examples.append(InputExample(guid=guid, text_a=text, text_b=None, label=None))
self.features = glue_convert_examples_to_features(examples, tokenizer,
max_length=max_seq_len,
label_list=label_list)
def __len__(self):
return len(self.features)
def __getitem__(self, i) -> InputFeatures:
return self.features[i]
def replace_ent_label(text, ent_type, start_idx, end_idx):
label = '@'+ent_type+'$'
return text[:start_idx]+label+text[end_idx:]
def write_file(file, index, sentence, label, sep, is_test, is_label):
if is_test and is_label: # test_original - test with labels
file.write('{}{}{}{}{}'.format(index, sep, sentence, sep, label))
elif is_test and not is_label: # test - test with no labels
file.write('{}{}{}'.format(index, sep, sentence))
else: # train
file.write('{}{}{}'.format(sentence, sep, label))
file.write('\n')
def get_char_split_points(record, max_len):
char_split_points = []
split_points = record.get_split_points(max_len=max_len)
for pt in split_points[:-1]:
char_split_points.append(record.get_char_idx(pt)[1])
if len(char_split_points) == 1:
return char_split_points
else:
return char_split_points[1:]
def replace_entity_text(split_text, ent1, ent2, split_offset):
# Remove split offset
ent1_start = ent1.range[0] - split_offset
ent1_end = ent1.range[1] - split_offset
ent2_start = ent2.range[0] - split_offset
ent2_end = ent2.range[1] - split_offset
# If entity 1 is present before entity 2
if ent1_end < ent2_end:
# Insert entity 2 and then entity 1
modified_text = replace_ent_label(split_text, ent2.name, ent2_start, ent2_end)
modified_text = replace_ent_label(modified_text, ent1.name, ent1_start, ent1_end)
# If entity 1 is present after entity 2
else:
# Insert entity 1 and then entity 2
modified_text = replace_ent_label(split_text, ent1.name, ent1_start, ent1_end)
modified_text = replace_ent_label(modified_text, ent2.name, ent2_start, ent2_end)
return modified_text
def generate_re_input_files(ehr_records: List[HealthRecord], filename: str,
ade_records: List[Dict] = None, max_len: int = 128,
is_test=False, is_label=True, is_predict=False, sep: str = '\t'):
random.seed(0)
index = 0
index_rel_label_map = []
with open(filename, 'w') as file:
# Write headers
write_file(file, 'index', 'sentence', 'label', sep, is_test, is_label)
# Preprocess EHR records
for record in ehr_records:
text = record.text
entities = record.get_entities()
if is_predict:
true_relations = None
else:
true_relations = record.get_relations()
# get character split points
char_split_points = get_char_split_points(record, max_len)
start = 0
end = char_split_points[0]
for i in range(len(char_split_points)):
# Obtain only entities within the split text
range_entities = {ent_id: ent for ent_id, ent in
filter(lambda item: int(item[1][0]) >= start and int(item[1][1]) <= end,
entities.items())}
# Get all possible relations within the split text
possible_relations = utils.map_entities(range_entities, true_relations)
for rel, label in possible_relations:
if label == 0 and rel.name != "ADE-Drug":
if random.random() > 0.25:
continue
split_text = text[start:end]
split_offset = start
ent1 = rel.get_entities()[0]
ent2 = rel.get_entities()[1]
# Check if both entities are within split text
if ent1.range[0] >= start and ent1.range[1] < end and \
ent2.range[0] >= start and ent2.range[1] < end:
modified_text = replace_entity_text(split_text, ent1, ent2, split_offset)
# Replace un-required characters with space
final_text = modified_text.replace('\n', ' ').replace('\t', ' ')
write_file(file, index, final_text, label, sep, is_test, is_label)
if is_predict:
index_rel_label_map.append({'relation': rel})
else:
index_rel_label_map.append({'label': label, 'relation': rel})
index += 1
start = end
if i != len(char_split_points)-1:
end = char_split_points[i+1]
else:
end = len(text)+1
# Preprocess ADE records
if ade_records is not None:
for record in ade_records:
entities = record['entities']
true_relations = record['relations']
possible_relations = utils.map_entities(entities, true_relations)
for rel, label in possible_relations:
if label == 1 and random.random() > 0.5:
continue
new_tokens = record['tokens'].copy()
for ent in rel.get_entities():
ent_type = ent.name
start_tok = ent.range[0]
end_tok = ent.range[1]+1
for i in range(start_tok, end_tok):
new_tokens[i] = '@'+ent_type+'$'
"""Remove consecutive repeating entities.
Eg. this is @ADE$ @ADE$ @ADE$ for @Drug$ @Drug$ -> this is @ADE$ for @Drug$"""
final_tokens = [new_tokens[i] for i in range(len(new_tokens))\
if (i == 0) or new_tokens[i] != new_tokens[i-1]]
final_text = " ".join(final_tokens)
write_file(file, index, final_text, label, sep, is_test, is_label)
index_rel_label_map.append({'label': label, 'relation': rel})
index += 1
filename, ext = filename.split('.')
utils.save_pickle(filename+'_rel.pkl', index_rel_label_map)
def get_eval_results(answer_path, output_path):
"""
Get evaluation metrics for predictions
Parameters
------------
answer_path : test.tsv file. Tab-separated.
One example per a line. True labels at the 3rd column.
output_path : test_predictions.txt. Model generated predictions.
"""
testdf = pd.read_csv(answer_path, sep="\t", index_col=0)
preddf = pd.read_csv(output_path, sep="\t", header=None)
pred = [preddf.iloc[i].tolist() for i in preddf.index]
pred_class = [int(v[1]) for v in pred[1:]]
p, r, f, s = precision_recall_fscore_support(y_pred=pred_class, y_true=testdf["label"])
results = dict()
results["f1 score"] = f[1]
results["recall"] = r[1]
results["precision"] = p[1]
results["specificity"] = r[0]
return results
def generate_re_test_file(ehr_record: HealthRecord,
max_len: int = 128) -> Tuple[List[str], List[Relation]]:
"""
Generates test file for Relation Extraction.
Parameters
-----------
ehr_record : HealthRecord
The EHR record with entities set.
max_len : int
The maximum length of sequence.
Returns
--------
Tuple[List[str], List[Relation]]
List of sequences with entity replaced by it's tag.
And a list of relation objects representing relation in those sequences.
"""
random.seed(0)
re_text_list = []
relation_list = []
text = ehr_record.text
entities = ehr_record.get_entities()
if isinstance(entities, dict):
entities = list(entities.values())
# get character split points
char_split_points = get_char_split_points(ehr_record, max_len)
start = 0
end = char_split_points[0]
for i in range(len(char_split_points)):
# Obtain only entities within the split text
range_entities = [ent for ent in filter(lambda item: int(item[0]) >= start and int(item[1]) <= end,
entities)]
# Get all possible relations within the split text
possible_relations = utils.map_entities(range_entities)
for rel, label in possible_relations:
split_text = text[start:end]
split_offset = start
ent1 = rel.get_entities()[0]
ent2 = rel.get_entities()[1]
# Check if both entities are within split text
if ent1[0] >= start and ent1[1] < end and \
ent2[0] >= start and ent2[1] < end:
modified_text = replace_entity_text(split_text, ent1, ent2, split_offset)
# Replace un-required characters with space
final_text = modified_text.replace('\n', ' ').replace('\t', ' ')
re_text_list.append(final_text)
relation_list.append(rel)
start = end
if i != len(char_split_points)-1:
end = char_split_points[i+1]
else:
end = len(text)+1
assert len(re_text_list) == len(relation_list)
return re_text_list, relation_list
