def test_spans_overlapping_tokens():
text = 'This is a sentence.'
tokens = [(0, 4), (5, 7), (8, 9), (10, 18), (18, 19)]
to = TokenOverlap(text, tokens)
assert to.overlapping_tokens(6, 11) == [(5, 7), (8, 9), (10, 18)]
assert to.overlapping_tokens(5, 15) == [(5, 7), (8, 9), (10, 18)]
def test_empty_overlap_between_tokens():
text = 'This is a sentence.'
tokens = [(0, 4), (5, 7), (8, 9), (10, 18), (18, 19)]
to = TokenOverlap(text, tokens)
assert len(to.overlapping_tokens(4, 5)) == 0
assert len(to.overlapping_tokens(7, 8)) == 0
assert len(to.overlapping_tokens(18, 18)) == 0
def test_empty_overlap_at_beginning():
text = ' This is a sentence.'
tokens = [(3, 7), (8, 10), (11, 12), (13, 21), (21, 22)]
to = TokenOverlap(text, tokens)
assert len(to.overlapping_tokens(0, 2)) == 0
assert len(to.overlapping_tokens(1, 3)) == 0
assert len(to.overlapping_tokens(0, 3)) == 0
def test_identical_spans():
text = 'This is a sentence.'
tokens = [(0, 4), (5, 7), (8, 9), (10, 18), (18, 19)]
to = TokenOverlap(text, tokens)
for start, end in tokens:
overlap = to.overlapping_tokens(start, end)
assert len(overlap) == 1
assert overlap[0] == (start, end)
def _compute_tp_total(self, input_gen):
exp_total = []
pred_total = []
for doc_index, document in enumerate(input_gen()):
assert doc_index < len(
self._documents
), 'Input generator yields more documents than expected!'
to = None
if self._token_func:
text = read(document.txt_path)
tokens = list(self._token_func(text))
to = TokenOverlap(text, tokens)
for anno_file_1, anno_file_2 in combinations(
document.ann_files, 2):
# tp, exp, pred, exp_list, pred_list = self._eval_func(anno_file_1.ann_path, anno_file_2.ann_path, tokens=to)
tp, exp, pred = self._eval_func(anno_file_1.ann_path,
anno_file_2.ann_path,
tokens=to)
# exp_total.append(exp_list)
# pred_total.append(pred_list)
pair_idx = self._pair2idx[(anno_file_1.annotator_id,
anno_file_2.annotator_id)]
doc_idx = self._doc2idx[document.doc_id]
self._increment_counts(tp, pair_idx, doc_idx, 0)
self._increment_counts(exp, pair_idx, doc_idx, 1)
self._increment_counts(pred, pair_idx, doc_idx, 1)
def find_all_matches_and_mismatches(self, input_gen):
# Finds all matches / mismatches (or, by other words: agreements & disagreements)
results = []
for doc_index, document in enumerate( input_gen() ):
assert doc_index < len(self._documents), 'Input generator yields more documents than expected!'
to = None
text = read(document.txt_path)
if self._token_func:
tokens = list(self._token_func(text))
to = TokenOverlap(text, tokens)
# TODO: the following solution robustly works for 2 annotators.
# in case of more than 2 annotators, it would be very nice
# to further consolidate the results
for anno_file_1, anno_file_2 in combinations(document.ann_files, 2):
tp, exp, pred = self._eval_func(anno_file_1.ann_path, anno_file_2.ann_path, tokens=to)
# Convert exp & pred to fp & fn
if self._token_func is not None:
fp = counter2list(Counter(pred) - Counter(exp))
fn = counter2list(Counter(exp) - Counter(pred))
else:
fp = pred.difference(exp)
fn = exp.difference(pred)
doc_result = {}
# all kinds of metadata
doc_result['doc_id'] = doc_index
doc_result['text_file'] = os.path.basename(document.txt_path)
doc_result['ann_file'] = anno_file_1.ann_path.name
doc_result['annotator_1'] = anno_file_1.annotator_id
doc_result['annotator_2'] = anno_file_2.annotator_id
# true-positive tuples: matching annotations
tp_tuples = self._extract_sorted_annotation_text_tuples(tp, text)
tp_tuples = self._format_text_tuples_for_pretty_printing(tp_tuples, skip_str_conversion=[0,1])
doc_result['matches'] = tp_tuples
# marked by annotator2, but not by annotator1
fp_tuples = self._extract_sorted_annotation_text_tuples(fp, text, annotator=anno_file_2.annotator_id)
# marked by annotator1, but not by annotator2
fn_tuples = self._extract_sorted_annotation_text_tuples(fn, text, annotator=anno_file_1.annotator_id)
all_mismatches = sorted(fp_tuples + fn_tuples, key=lambda x:x[2])
all_mismatches = self._format_text_tuples_for_pretty_printing(all_mismatches, skip_str_conversion=[0,1,2])
doc_result['mismatches'] = all_mismatches
results.append( doc_result )
return results