def rel_emb_score(premise, hypothesis, type_prem, id_prem, type_hypo, id_hypo,
is_premise_reversed, is_hypothesis_reversed):
global rel_emb_model
if rel_emb_model is None:
logger.info("Loading relation embeddings.")
rel_emb_model = KeyedVectors.load_word2vec_format(
"embeddings/filtered/typed_rel_emb.txt", binary=False)
logger.info("Done.")
return rel_emb_model.similarity(combine_type(type_prem, id_prem),
combine_type(type_hypo, id_hypo))
def extract_relations(datasets: List[str], with_types: bool):
rels = set()
for d in datasets:
for row in read_in_dataset(d):
if with_types:
pr = combine_type(row[2], row[3])
hy = combine_type(row[4], row[5])
else:
pr = str(row[3])
hy = str(row[5])
rels.add(pr)
rels.add(hy)
return rels
def complEx_typed_score(premise, hypothesis, type_prem, id_prem, type_hypo,
id_hypo, is_premise_reversed, is_hypothesis_reversed):
global typed_ComplEx_re_model
global typed_ComplEx_im_model
if typed_ComplEx_re_model is None:
logger.info("Loading typed ComplEx embeddings.")
typed_ComplEx_re_model = KeyedVectors.load_word2vec_format(
"embeddings/filtered/typed_ComplEx-real.txt", binary=False)
typed_ComplEx_im_model = KeyedVectors.load_word2vec_format(
"embeddings/filtered/typed_ComplEx-imaginary.txt", binary=False)
logger.info("Done.")
prem_vec = build_complex_vector(combine_type(type_prem, id_prem),
typed_ComplEx_re_model,
typed_ComplEx_im_model)
hypo_vec = build_complex_vector(combine_type(type_hypo, id_hypo),
typed_ComplEx_re_model,
typed_ComplEx_im_model)
return np.real(
prem_vec.dot(hypo_vec) / np.linalg.norm(prem_vec) /
np.linalg.norm(hypo_vec))
def get_vectors(id1, type1, id2, type2, typed=True):
if typed:
id1, id2 = combine_type(type1, id1), combine_type(type2, id2)
return vecForRel(id1, typed=typed), vecForRel(id2, typed=typed)