def mask_examples(nlp: Pipeline, domain: str, is_test: bool):
dataset = absa.load_examples('semeval', domain, is_test)
for i, example in enumerate(dataset):
yield i, -1, example # Predict without a mask.
[tokenized_example] = nlp.tokenize([example])
n = len(tokenized_example.text_tokens)
for index in range(n):
new_example = mask_tokens(nlp, tokenized_example, indices=[index])
yield i, index, new_example
def test_predict(nlp: Pipeline):
text_1 = ("We are great fans of Slack, but we wish the subscriptions "
"were more accessible to small startups.")
text_2 = "We are great fans of Slack"
aspect = "Slack"
examples = [Example(text_1, aspect), Example(text_2, aspect)]
tokenized_examples = nlp.tokenize(examples)
input_batch = nlp.encode(tokenized_examples)
output_batch = nlp.predict(input_batch)
assert output_batch.scores.shape == [2, 3]
assert output_batch.hidden_states.shape == [2, 13, 25, 768]
assert output_batch.attentions.shape == [2, 12, 12, 25, 25]
assert output_batch.attention_grads.shape == [2, 12, 12, 25, 25]
scores = output_batch.scores.numpy()
assert np.argmax(scores, axis=-1).tolist() == [2, 2]
def mask_examples(nlp: Pipeline, domain: str, part_parts: Tuple[int, int]):
dataset = absa.load_examples('semeval', domain, test=True)
# Split a dataset because it's better to cache more freq.
part, parts = part_parts
chunks = utils.split(dataset, n=parts)
dataset_chunk = chunks[part]
for i, example in enumerate(dataset_chunk):
yield i, -1, -1, example # Predict without a mask.
[tokenized_example] = nlp.tokenize([example])
n = len(tokenized_example.text_tokens)
x, y = np.triu_indices(n, k=1) # Exclude the diagonal.
for ij in zip(x, y):
new_example = mask_tokens(nlp, tokenized_example, indices=ij)
yield i, *ij, new_example
def test_review(nlp: Pipeline):
text_1 = ("We are great fans of Slack, but we wish the subscriptions "
"were more accessible to small startups.")
text_2 = "The Slack often has bugs."
aspect = "slack"
examples = [Example(text_1, aspect), Example(text_2, aspect)]
tokenized_examples = nlp.tokenize(examples)
input_batch = nlp.encode(tokenized_examples)
output_batch = nlp.predict(input_batch)
predictions = nlp.review(tokenized_examples, output_batch)
predictions = list(predictions)
labeled_1, labeled_2 = predictions
assert labeled_1.sentiment == Sentiment.positive
assert labeled_2.sentiment == Sentiment.negative
assert isinstance(labeled_1, PredictedExample)
assert isinstance(labeled_1.scores, list)
assert isinstance(labeled_1.review, Review)
assert not labeled_1.review.is_reference
assert not labeled_1.review.patterns
def test_encode(nlp: Pipeline):
text_1 = ("We are great fans of Slack, but we wish the subscriptions "
"were more accessible to small startups.")
text_2 = "We are great fans of Slack"
aspect = "Slack"
examples = [Example(text_1, aspect), Example(text_2, aspect)]
tokenized_examples = nlp.tokenize(examples)
input_batch = nlp.encode(tokenized_examples)
assert isinstance(input_batch.token_ids, tf.Tensor)
# 101 the CLS token, 102 the SEP tokens.
token_ids = input_batch.token_ids.numpy()
values = [101, 2057, 2024, 2307, 4599, 1997, 19840, 102, 19840, 102]
assert token_ids[1, :10].tolist() == values
assert token_ids[0, :7].tolist() == values[:7]
# The second sequence should be padded (shorter),
# and attention mask should be set.
assert np.allclose(token_ids[1, 10:], 0)
attention_mask = input_batch.attention_mask.numpy()
assert np.allclose(attention_mask[1, 10:], 0)
# Check how the tokenizer marked the segments.
token_type_ids = input_batch.token_type_ids.numpy()
assert token_type_ids[0, -2:].tolist() == [1, 1]
assert np.allclose(token_type_ids[0, :-2], 0)
def mask_examples(
nlp: Pipeline,
domain: str,
part_parts: Tuple[int, int]
):
dataset = absa.load_examples('semeval', domain, test=True)
# Filter out examples that contain a key token or a pair of key tokens,
# and that are other than negative.
y_ref, _, mask_1 = key_token_labels(nlp, domain, is_test=True)
y_ref, _, mask_2 = key_token_pair_labels(nlp, domain, parts=10)
mask = ~(mask_1 | mask_2) & y_ref == Sentiment.negative.value
dataset = [e for e, is_correct in zip(dataset, mask) if is_correct]
# Split a dataset because it's better to cache more freq.
part, parts = part_parts
chunks = utils.split(dataset, n=parts)
dataset_chunk = chunks[part]
for i, example in enumerate(dataset_chunk):
yield i, -1, -1, -1, example # Predict without a mask.
[tokenized_example] = nlp.tokenize([example])
n = len(tokenized_example.text_tokens)
ij = np.zeros(shape=[n, n])
ij[np.triu_indices(n, k=1)] = 1 # The j shifted by 1 from i.
ij = ij.reshape([n, n, 1]).astype(bool)
jk = np.zeros(shape=[n, n])
jk[np.triu_indices(n, k=1)] = 1 # The k shifted by 1 from j.
jk = jk.reshape([1, n, n]).astype(bool)
matrix_ijk = np.where(ij & jk)
for ijk in zip(*matrix_ijk):
new_example = mask_tokens(nlp, tokenized_example, indices=ijk)
yield i, *ijk, new_example