qasper / dataset_reader.py
ag2435's picture
added original qasper data & preprocessed version using dataset_reader.py script
a8de101
"""
Adapted from https://github.com/allenai/qasper-led-baseline/blob/main/qasper_baselines/dataset_reader.py
to get ride of allennlp dependencies.
"""
import json
import logging
import random
from enum import Enum
from collections import defaultdict
from typing import Any, Dict, List, Optional, Iterable, Tuple
# from overrides import overrides
# import spacy
import torch
# from allennlp.common.util import JsonDict
# from allennlp.data.fields import (
# MetadataField,
# TextField,
# IndexField,
# ListField,
# TensorField,
# )
# from allennlp.common.file_utils import cached_path, open_compressed
# from allennlp.data.dataset_readers.dataset_reader import DatasetReader
# from allennlp.data.instance import Instance
# from allennlp.data.token_indexers import PretrainedTransformerIndexer
# from allennlp.data.tokenizers import Token, PretrainedTransformerTokenizer
from transformers import AutoTokenizer
logger = logging.getLogger(__name__)
class AnswerType(Enum):
EXTRACTIVE = 1
ABSTRACTIVE = 2
BOOLEAN = 3
NONE = 4
# @DatasetReader.register("qasper")
class QasperReader(object):
"""
Reads a JSON-formatted Qasper data file and returns a `Dataset` where the `Instances` have
four fields:
* `question_with_context`, a `TextField` that contains the concatenation of question and
context,
* `paragraph_indices`, a `ListField` of `IndexFields` indicating paragraph-start tokens
in `question_with_context`.
* `global_attention_mask`, a mask that can be used by a longformer to specify which tokens in
`question_with_context` should have global attention (only present if
`include_global_attention_mask` is `True`).
* `evidence`, a 0/1 `TensorField` indicating whether each paragraph in `paragraph_indices`
should be selected as evidence.
* `answer`, a `TextField` that contains the (wordpiece-tokenized) answer to the question
* A `MetadataField` that stores the instance's ID, paper ID, the original question, the
original passage text, both of these in tokenized form, and the context also broken into
paragraphs, and the gold evidence spans, accessible as `metadata['question_id']`,
`metadata['article_id']`, `metadata['question']`, `metadata['context']`,
`metadata['question_tokens']`, `metadata['context_tokens']`,
`metadata['context_paragraphs']`, `metadata['all_evidence']`, `metadata['all_answers']`.
Parameters
----------
transformer_model_name : `str`, optional (default=`allenai/led-large-16384`)
This reader chooses tokenizer and token indexer according to this setting.
max_query_length : `int`, optional (default=128)
The maximum number of wordpieces dedicated to the question. If the question is longer than
this, it will be truncated.
max_document_length : `int` , optional (default=16384)
This is the maximum number of wordpieces allowed per one whole document (including the
question, for simplicity). If the document is longer than this many word pieces, it will be
truncated.
paragraph_separator : `Optional[str]`, optional (default="</s>")
If given, we will use this as a separator token in between paragraphs. Pass in `None` to
have this not be used.
include_global_attention_mask : `bool` (default = True)
If `True`, we will include a field in the output containing a global attention mask for use
with a longformer, which is `True` for all starts of paragraphs and question tokens, so
attention will always be placed on those tokens.
context : `str` (default = `full_text`)
To reproduce the baselines from the paper that do not have access to the full text of the paper
you can change this argument. Options are `question_only`, `question_and_abstract`,
`question_and_introduction`, `question_and_evidence`. If this is set to `question_andevidence`,
the reader will ignore answers that are `None`, and those that are boolean.
for_training : `bool` (default = False)
This flag affects how questions with multiple answers are handled. When set to True, this flag
causes the reader to yield one instance per answer. When set to False, the instance will contain
only the first answer. The metadata will always contain all the answers and evidence, which can be
used at evaluation time to compute aggregated metrics.
"""
def __init__(
self,
transformer_model_name: str = "allenai/led-base-16384",
max_query_length: int = 128,
max_document_length: int = 16384,
paragraph_separator: Optional[str] = "</s>",
include_global_attention_mask: bool = True,
context: str = "full_text",
for_training: bool = False,
**kwargs,
) -> None:
# super().__init__(
# manual_distributed_sharding=True,
# manual_multiprocess_sharding=True,
# **kwargs,
# )
self._transformer_model_name = transformer_model_name
# self._tokenizer = PretrainedTransformerTokenizer(
# transformer_model_name, add_special_tokens=False
# )
self._tokenizer = AutoTokenizer.from_pretrained(transformer_model_name)
# Albert: hack
self._tokenizer.sequence_pair_start_tokens = [self._tokenizer.bos_token,]
self._include_global_attention_mask = include_global_attention_mask
# self._token_indexers = {
# "tokens": PretrainedTransformerIndexer(transformer_model_name)
# }
self.max_query_length = max_query_length
self.max_document_length = max_document_length
self._paragraph_separator = paragraph_separator
if context not in [
"full_text",
"question_only",
"question_and_abstract",
"question_and_introduction",
"question_and_evidence"
]:
raise RuntimeError(f"Unrecognized context type: {context}")
self._context = context
self._for_training = for_training
self._stats = defaultdict(int)
# @overrides
def _read(self, file_path: str):
# if `file_path` is a URL, redirect to the cache
# file_path = cached_path(file_path)
logger.info("Reading the dataset")
if file_path.endswith(".json"):
yield from self._read_json(file_path)
elif file_path.endswith(".jsonl"):
yield from self._read_json_lines(file_path)
else:
raise RuntimeError(
f"Unsupported extension on file: {file_path}. Only json and jsonl are supported."
)
def _read_json(self, file_path: str):
logger.info("Reading json file at %s", file_path)
with open(file_path, 'r') as dataset_file:
dataset = json.load(dataset_file)
for article_id, article in dataset.items():
if not article["full_text"]:
continue
article["article_id"] = article_id
yield from self._article_to_instances(article)
self._log_stats()
def _read_json_lines(self, file_path: str):
logger.info("Reading json lines file at %s", file_path)
with open(file_path, 'r') as dataset_file:
for data_line in self.shard_iterable(dataset_file):
data = json.loads(data_line)
yield from self._article_to_instances(data)
self._log_stats()
def _log_stats(self) -> None:
logger.info("Stats:")
for key, value in self._stats.items():
logger.info("%s: %d", key, value)
def _article_to_instances(self, article: Dict[str, Any]):
paragraphs = self._get_paragraphs_from_article(article)
tokenized_context = None
paragraph_start_indices = None
# If the context is evidence, text_to_instance will make the appropriate tokenized_context.
if not self._context == "question_and_evidence":
tokenized_context, paragraph_start_indices = self._tokenize_paragraphs(
paragraphs
)
self._stats["number of documents"] += 1
for question_answer in article["qas"]:
self._stats["number of questions"] += 1
self._stats["number of answers"] += len(question_answer["answers"])
if len(question_answer["answers"]) > 1:
self._stats["questions with multiple answers"] += 1
all_answers = []
all_evidence = []
all_evidence_masks = []
for answer_annotation in question_answer["answers"]:
answer, evidence, answer_type = self._extract_answer_and_evidence(
answer_annotation["answer"]
)
all_answers.append({"text": answer, "type": answer_type})
all_evidence.append(evidence)
evidence_mask = self._get_evidence_mask(evidence, paragraphs)
all_evidence_masks.append(evidence_mask)
additional_metadata = {
"question_id": question_answer["question_id"],
"article_id": article.get("article_id"),
"all_answers": all_answers,
"all_evidence": all_evidence,
"all_evidence_masks": all_evidence_masks,
}
answers_to_yield = [x['text'] for x in all_answers] if self._for_training else [all_answers[0]['text']]
evidence_masks_to_yield = all_evidence_masks if self._for_training else [all_evidence_masks[0]]
evidence_to_yield = all_evidence if self._for_training else [all_evidence[0]]
for answer, evidence, evidence_mask in zip(answers_to_yield, evidence_to_yield, evidence_masks_to_yield):
if self._context == "question_and_evidence" and answer in ['Unanswerable', 'Yes', 'No']:
continue
yield self.text_to_instance(
question_answer["question"],
paragraphs,
tokenized_context,
paragraph_start_indices,
evidence_mask,
answer,
evidence,
additional_metadata,
)
@staticmethod
def _get_evidence_mask(evidence: List[str], paragraphs: List[str]) -> List[int]:
"""
Takes a list of evidence snippets, and the list of all the paragraphs from the
paper, and returns a list of indices of the paragraphs that contain the evidence.
"""
evidence_mask = []
for paragraph in paragraphs:
for evidence_str in evidence:
if evidence_str in paragraph:
evidence_mask.append(1)
break
else:
evidence_mask.append(0)
return evidence_mask
# @overrides
def text_to_instance(
self, # type: ignore # pylint: disable=arguments-differ
question: str,
paragraphs: List[str],
tokenized_context: List = None,
paragraph_start_indices: List[int] = None,
evidence_mask: List[int] = None,
answer: str = None,
evidence: List[str] = None,
additional_metadata: Dict[str, Any] = None):
fields = {}
tokenized_question = self._tokenizer.tokenize(question)
if len(tokenized_question) > self.max_query_length:
self._stats["number of truncated questions"] += 1
tokenized_question = tokenized_question[:self.max_query_length]
if tokenized_context is None or paragraph_start_indices is None:
if self._context == "question_and_evidence":
tokenized_context, paragraph_start_indices = self._tokenize_paragraphs(
evidence
)
else:
tokenized_context, paragraph_start_indices = self._tokenize_paragraphs(
paragraphs
)
allowed_context_length = (
self.max_document_length
- len(tokenized_question)
- len(self._tokenizer.sequence_pair_start_tokens)
- 1 # for paragraph seperator
)
if len(tokenized_context) > allowed_context_length:
self._stats["number of truncated contexts"] += 1
tokenized_context = tokenized_context[:allowed_context_length]
paragraph_start_indices = [index for index in paragraph_start_indices
if index <= allowed_context_length]
if evidence_mask is not None:
num_paragraphs = len(paragraph_start_indices)
evidence_mask = evidence_mask[:num_paragraphs]
# This is what Iz's code does.
question_and_context = (
self._tokenizer.sequence_pair_start_tokens
+ tokenized_question
+ [self._paragraph_separator]
+ tokenized_context
)
# make the question field
question_field = question_and_context
fields["question_with_context"] = question_field
start_of_context = (
len(self._tokenizer.sequence_pair_start_tokens)
+ len(tokenized_question)
)
paragraph_indices_list = [x + start_of_context for x in paragraph_start_indices]
paragraph_indices_field = (
[x for x in paragraph_indices_list] if paragraph_indices_list else
[-1]
)
fields["paragraph_indices"] = paragraph_indices_field
if self._include_global_attention_mask:
# We need to make a global attention array. We'll use all the paragraph indices and the
# indices of question tokens.
mask_indices = set(list(range(start_of_context)) + paragraph_indices_list)
mask = [
True if i in mask_indices else False for i in range(len(question_field))
]
fields["global_attention_mask"] = torch.tensor(mask)
if evidence_mask is not None:
# evidence_field = torch.tensor(evidence_mask)
evidence_field = evidence_mask
fields["evidence"] = evidence_field
if answer:
# fields["answer"] = (
# self._tokenizer.add_special_tokens(self._tokenizer.tokenize(answer))
# )
fields["answer"] = self._tokenizer.tokenize(answer) #, add_special_tokens=True)
# make the metadata
metadata = {
"question": question,
"question_tokens": tokenized_question,
"paragraphs": paragraphs,
"context_tokens": tokenized_context,
}
if additional_metadata is not None:
metadata.update(additional_metadata)
fields["metadata"] = metadata
return fields
# @overrides
def apply_token_indexers(self, instance) -> None:
instance.fields["question_with_context"].token_indexers = self._token_indexers
instance.fields["answer"].token_indexers = self._token_indexers
def _tokenize_paragraphs(
self, paragraphs: List[str]):
tokenized_context = []
paragraph_start_indices = []
for paragraph in paragraphs:
tokenized_paragraph = self._tokenizer.tokenize(paragraph)
paragraph_start_indices.append(len(tokenized_context))
tokenized_context.extend(tokenized_paragraph)
if self._paragraph_separator:
tokenized_context.append(self._paragraph_separator)
if self._paragraph_separator:
# We added the separator after every paragraph, so we remove it after the last one.
tokenized_context = tokenized_context[:-1]
return tokenized_context, paragraph_start_indices
def _extract_answer_and_evidence(
self, answer: List
) -> Tuple[str, List[str]]:
evidence_spans = [x.replace("\n", " ").strip() for x in answer["evidence"]]
evidence_spans = [x for x in evidence_spans if x != ""]
if not evidence_spans:
self._stats["answers with no evidence"] += 1
# TODO (pradeep): Deal with figures and tables.
if any(["FLOAT SELECTED" in span for span in evidence_spans]):
# Ignoring question if any of the selected evidence is a table or a figure.
self._stats["answers with table or figure as evidence"] += 1
if len(evidence_spans) > 1:
self._stats["multiple_evidence_spans_count"] += 1
answer_string = None
answer_type = None
if answer.get("unanswerable", False):
self._stats["unanswerable questions"] += 1
answer_string = "Unanswerable"
answer_type = AnswerType.NONE.name
elif answer.get("yes_no") is not None:
self._stats["yes/no questions"] += 1
answer_string = "Yes" if answer["yes_no"] else "No"
answer_type = AnswerType.BOOLEAN.name
elif answer.get("extractive_spans", []):
self._stats["extractive questions"] += 1
if len(answer["extractive_spans"]) > 1:
self._stats["extractive questions with multiple spans"] += 1
answer_string = ", ".join(answer["extractive_spans"])
answer_type = AnswerType.EXTRACTIVE.name
else:
answer_string = answer.get("free_form_answer", "")
if not answer_string:
self._stats["questions with empty answer"] += 1
else:
self._stats["freeform answers"] += 1
answer_type = AnswerType.ABSTRACTIVE.name
return answer_string, evidence_spans, answer_type
def _get_paragraphs_from_article(self, article: Dict) -> List[str]:
if self._context == "question_only":
return []
if self._context == "question_and_abstract":
return [article["abstract"]]
full_text = article["full_text"]
paragraphs = []
for section_info in full_text:
# TODO (pradeep): It is possible there are other discrepancies between plain text, LaTeX and HTML.
# Do a thorough investigation and add tests.
if section_info["section_name"] is not None:
paragraphs.append(section_info["section_name"])
for paragraph in section_info["paragraphs"]:
paragraph_text = paragraph.replace("\n", " ").strip()
if paragraph_text:
paragraphs.append(paragraph_text)
if self._context == "question_and_introduction":
# Assuming the first section is the introduction and stopping here.
break
return paragraphs