def average_cosine(complex_sentence, simple_sentence):
complex_embeddings = to_embeddings(complex_sentence)
simple_embeddings = to_embeddings(simple_sentence)
return float(
F.cosine_similarity(complex_embeddings.mean(dim=0),
simple_embeddings.mean(dim=0),
dim=0))
def hungarian_dot(complex_sentence, simple_sentence):
complex_embeddings = to_embeddings(complex_sentence)
simple_embeddings = to_embeddings(simple_sentence)
similarity_matrix = torch.mm(complex_embeddings, simple_embeddings.t())
row_indexes, col_indexes = linear_sum_assignment(-similarity_matrix)
# TODO: Penalize less deletion of unimportant words
return float(similarity_matrix[row_indexes, col_indexes].sum() / max(len(complex_sentence), len(simple_sentence)))
def hungarian_cosine(complex_sentence, simple_sentence):
complex_embeddings = to_embeddings(complex_sentence)
simple_embeddings = to_embeddings(simple_sentence)
similarity_matrix = torch.zeros(len(complex_embeddings), len(simple_embeddings))
for (i, complex_embedding), (j, simple_embedding) in itertools.product(enumerate(complex_embeddings),
enumerate(simple_embeddings)):
similarity_matrix[i, j] = F.cosine_similarity(complex_embedding, simple_embedding, dim=0)
row_indexes, col_indexes = linear_sum_assignment(-similarity_matrix)
# TODO: Penalize less deletion of unimportant words
return float(similarity_matrix[row_indexes, col_indexes].sum() / max(len(complex_sentence), len(simple_sentence)))
def average_dot(complex_sentence, simple_sentence):
complex_embeddings = to_embeddings(complex_sentence)
simple_embeddings = to_embeddings(simple_sentence)
return float(
torch.dot(complex_embeddings.mean(dim=0),
simple_embeddings.mean(dim=0)))
