def heuristic_fn(text, candidate):
'''
Return the difference between the classification probability of the original
word and the candidate substitute synonym, which is defined in Eq.(4) and Eq.(5).
'''
doc = nlp(text)
origin_vector = None
perturbed_vector = None
if level == 'word':
origin_vector = text_to_vector(text, tokenizer, dataset)
perturbed_tokens = _compile_perturbed_tokens(doc, [candidate])
perturbed_doc = nlp(' '.join(perturbed_tokens))
perturbed_vector = text_to_vector(perturbed_doc.text, tokenizer,
dataset)
elif level == 'char':
max_len = config.char_max_len[dataset]
origin_vector = doc_process(text, get_embedding_dict(),
dataset).reshape(1, max_len)
perturbed_tokens = _compile_perturbed_tokens(
nlp(input_text), [candidate])
perturbed_text = ' '.join(perturbed_tokens)
perturbed_vector = doc_process(perturbed_text,
get_embedding_dict(),
dataset).reshape(1, max_len)
origin_prob = grad_guide.predict_prob(input_vector=origin_vector)
perturbed_prob = grad_guide.predict_prob(input_vector=perturbed_vector)
delta_p = origin_prob[true_y] - perturbed_prob[true_y]
return delta_p
def heuristic_fn(text, candidate):
'''
Return the difference between the forward derivative of the original
word and the candidate substitute synonym, amplified by synonym
relevance rank.
Yes, this one is pretty bad in terms of performance.
'''
doc = nlp(text)
x = extract_features([doc], max_length=max_length)[0] \
.reshape(1, -1)
grads = grad_guide.wordwise_grads(x).squeeze()
index = candidate.token_position
derivative = grads[index]
perturbed_tokens = _compile_perturbed_tokens(doc, [candidate])
perturbed_doc = nlp(' '.join(perturbed_tokens))
perturbed_x = extract_features(
[perturbed_doc], max_length=max_length)[0] \
.reshape(1, -1)
perturbed_grads = grad_guide.wordwise_grads(perturbed_x).squeeze()
perturbed_derivative = perturbed_grads[index]
rank = candidate.similarity_rank + 1
raw_score = derivative - perturbed_derivative
raw_score *= -1 * target
return raw_score / rank