def _read(self, file_path):
data_dir = os.path.dirname(file_path)
annotations = annotations_from_jsonl(file_path)
documents: Dict[str, List[str]] = load_flattened_documents(data_dir,
docids=None)
for _, line in enumerate(annotations):
annotation_id: str = line.annotation_id
evidences: List[List[Evidence]] = line.evidences
label: str = line.classification
query: str = line.query
docids: List[str] = sorted(
list(
set([
evclause.docid for evgroup in evidences
for evclause in evgroup
])))
filtered_documents: Dict[str, List[str]] = dict([(d, documents[d])
for d in docids])
document_evidence_map = self.generate_document_evidence_map(
evidences)
if label is not None:
label = str(label)
instance = self.text_to_instance(
annotation_id=annotation_id,
documents=filtered_documents,
rationales=document_evidence_map,
query=query,
label=label,
)
if instance is not None:
yield instance
def main():
parser = argparse.ArgumentParser(description="""Computes rationale and final class classification scores""", formatter_class=argparse.RawTextHelpFormatter)
parser.add_argument('--data_dir', dest='data_dir', required=True, help='Which directory contains a {train,val,test}.jsonl file?')
parser.add_argument('--split', dest='split', required=True, help='Which of {train,val,test} are we scoring on?')
parser.add_argument('--strict', dest='strict', required=False, action='store_true', default=False, help='Do we perform strict scoring?')
parser.add_argument('--results', dest='results', required=True, help="""Results File
Contents are expected to be jsonl of:
{
"annotation_id": str, required
# these classifications *must not* overlap
# these classifications *must not* overlap
"rationales": List[
{
"docid": str, required
"hard_rationale_predictions": List[{
"start_token": int, inclusive, required
"end_token": int, exclusive, required
}], optional,
# token level classifications, a value must be provided per-token
# in an ideal world, these correspond to the hard-decoding above.
"soft_rationale_predictions": List[float], optional.
# sentence level classifications, a value must be provided for every
# sentence in each document, or not at all
"soft_sentence_predictions": List[float], optional.
}
],
# the classification the model made for the overall classification task
"classification": str, optional
# A probability distribution output by the model. We require this to be normalized.
"classification_scores": Dict[str, float], optional
# The next two fields are measures for how faithful your model is (the
# rationales it predicts are in some sense causal of the prediction), and
# how sufficient they are. We approximate a measure for comprehensiveness by
# asking that you remove the top k%% of tokens from your documents,
# running your models again, and reporting the score distribution in the
# "comprehensiveness_classification_scores" field.
# We approximate a measure of sufficiency by asking exactly the converse
# - that you provide model distributions on the removed k%% tokens.
# 'k' is determined by human rationales, and is documented in our paper.
# You should determine which of these tokens to remove based on some kind
# of information about your model: gradient based, attention based, other
# interpretability measures, etc.
# scores per class having removed k%% of the data, where k is determined by human comprehensive rationales
"comprehensiveness_classification_scores": Dict[str, float], optional
# scores per class having access to only k%% of the data, where k is determined by human comprehensive rationales
"sufficiency_classification_scores": Dict[str, float], optional
# the number of tokens required to flip the prediction - see "Is Attention Interpretable" by Serrano and Smith.
"tokens_to_flip": int, optional
}
When providing one of the optional fields, it must be provided for *every* instance.
The classification, classification_score, and comprehensiveness_classification_scores
must together be present for every instance or absent for every instance.
""")
parser.add_argument('--iou_thresholds', dest='iou_thresholds', required=False, nargs='+', type=float, default=[0.5], help='''Thresholds for IOU scoring.
These are used for "soft" or partial match scoring of rationale spans.
A span is considered a match if the size of the intersection of the prediction
and the annotation, divided by the union of the two spans, is larger than
the IOU threshold. This score can be computed for arbitrary thresholds.
''')
parser.add_argument('--score_file', dest='score_file', required=False, default=None, help='Where to write results?')
args = parser.parse_args()
results = load_jsonl(args.results)
docids = set(chain.from_iterable([rat['docid'] for rat in res['rationales']] for res in results))
docs = load_flattened_documents(args.data_dir, docids)
verify_instances(results, docs)
# load truth
annotations = annotations_from_jsonl(os.path.join(args.data_dir, args.split + '.jsonl'))
docids |= set(chain.from_iterable((ev.docid for ev in chain.from_iterable(ann.evidences)) for ann in annotations))
has_final_predictions = _has_classifications(results)
scores = dict()
if args.strict:
if not args.iou_thresholds:
raise ValueError("--iou_thresholds must be provided when running strict scoring")
if not has_final_predictions:
raise ValueError("We must have a 'classification', 'classification_score', and 'comprehensiveness_classification_score' field in order to perform scoring!")
# TODO think about offering a sentence level version of these scores.
if _has_hard_predictions(results):
truth = list(chain.from_iterable(Rationale.from_annotation(ann) for ann in annotations))
pred = list(chain.from_iterable(Rationale.from_instance(inst) for inst in results))
if args.iou_thresholds is not None:
iou_scores = partial_match_score(truth, pred, args.iou_thresholds)
scores['iou_scores'] = iou_scores
# NER style scoring
rationale_level_prf = score_hard_rationale_predictions(truth, pred)
scores['rationale_prf'] = rationale_level_prf
token_level_truth = list(chain.from_iterable(rat.to_token_level() for rat in truth))
token_level_pred = list(chain.from_iterable(rat.to_token_level() for rat in pred))
token_level_prf = score_hard_rationale_predictions(token_level_truth, token_level_pred)
scores['token_prf'] = token_level_prf
else:
logging.info("No hard predictions detected, skipping rationale scoring")
if _has_soft_predictions(results):
flattened_documents = load_flattened_documents(args.data_dir, docids)
paired_scoring = PositionScoredDocument.from_results(results, annotations, flattened_documents, use_tokens=True)
token_scores = score_soft_tokens(paired_scoring)
scores['token_soft_metrics'] = token_scores
else:
logging.info("No soft predictions detected, skipping rationale scoring")
if _has_soft_sentence_predictions(results):
documents = load_documents(args.data_dir, docids)
paired_scoring = PositionScoredDocument.from_results(results, annotations, documents, use_tokens=False)
sentence_scores = score_soft_tokens(paired_scoring)
scores['sentence_soft_metrics'] = sentence_scores
else:
logging.info("No sentence level predictions detected, skipping sentence-level diagnostic")
if has_final_predictions:
flattened_documents = load_flattened_documents(args.data_dir, docids)
class_results = score_classifications(results, annotations, flattened_documents)
scores['classification_scores'] = class_results
else:
logging.info("No classification scores detected, skipping classification")
pprint.pprint(scores)
if args.score_file:
with open(args.score_file, 'w') as of:
json.dump(scores, of, indent=4, sort_keys=True)