def test_elmo_offline():
mentions = get_embedd_mentions()
embeder = ElmoEmbedding()
for mention in mentions:
embeder.get_head_feature_vector(mention)
elmo_embeddings = WordEmbeddingRelationExtraction(EmbeddingMethod.ELMO_OFFLINE, elmo_file=None)
elmo_embeddings.embedding.embeder = embeder.cache
assert (
elmo_embeddings.extract_all_relations(mentions[0], mentions[0]).pop()
== RelationType.WORD_EMBEDDING_MATCH
)
assert (
elmo_embeddings.extract_all_relations(mentions[0], mentions[1]).pop()
== RelationType.WORD_EMBEDDING_MATCH
)
assert (
elmo_embeddings.extract_all_relations(mentions[0], mentions[2]).pop()
== RelationType.NO_RELATION_FOUND
)
assert (
elmo_embeddings.extract_all_relations(mentions[0], mentions[3]).pop()
== RelationType.WORD_EMBEDDING_MATCH
)
def load_elmo_for_vocab(vocabulary):
elmo_embeddings = ElmoEmbedding()
elmo_dict = dict()
for mention_string in vocabulary:
if mention_string not in elmo_dict:
mention_embedding = elmo_embeddings.get_avrg_feature_vector(
mention_string)
elmo_dict[mention_string] = mention_embedding
return elmo_dict
def __init__(
self,
method: EmbeddingMethod = EmbeddingMethod.GLOVE,
glove_file: str = None,
elmo_file: str = None,
cos_accepted_dist: float = 0.7,
):
"""
Extract Relation between two mentions according to Word Embedding cosine distance
Args:
method (optional): EmbeddingMethod.{GLOVE/GLOVE_OFFLINE/ELMO/ELMO_OFFLINE}
(default = GLOVE)
glove_file (required on GLOVE/GLOVE_OFFLINE mode): str Location of Glove file
elmo_file (required on ELMO_OFFLINE mode): str Location of Elmo file
"""
if method == EmbeddingMethod.GLOVE:
self.embedding = GloveEmbedding(glove_file)
self.contextual = False
elif method == EmbeddingMethod.GLOVE_OFFLINE:
self.embedding = GloveEmbeddingOffline(glove_file)
self.contextual = False
elif method == EmbeddingMethod.ELMO:
self.embedding = ElmoEmbedding()
self.contextual = True
elif method == EmbeddingMethod.ELMO_OFFLINE:
self.embedding = ElmoEmbeddingOffline(elmo_file)
self.contextual = True
self.accepted_dist = cos_accepted_dist
super(WordEmbeddingRelationExtraction, self).__init__()
def load_elmo_for_vocab(mentions):
"""
Create the embedding using the cache logic in the embedding class
Args:
mentions:
Returns:
"""
elmo_embeddings = ElmoEmbedding()
for mention in mentions:
elmo_embeddings.get_head_feature_vector(mention)
logger.info('Total words/contexts in vocabulary %d', len(elmo_embeddings.cache))
return elmo_embeddings.cache
def __init__(self, method: EmbeddingMethod, glove_file: str = None, elmo_file: str = None):
"""
Extract Relation between two mentions according to Word Embedding cosine distance
Args:
method (required): EmbeddingMethod.{GLOVE/GLOVE_OFFLINE/ELMO/ELMO_OFFLINE}
glove_file (required on GLOVE/GLOVE_OFFLINE mode): str Location of Glove file
elmo_file (required on ELMO_OFFLINE mode): str Location of Elmo file
"""
if method == EmbeddingMethod.GLOVE:
self.embedding = GloveEmbedding(glove_file)
elif method == EmbeddingMethod.GLOVE_OFFLINE:
self.embedding = GloveEmbeddingOffline(glove_file)
elif method == EmbeddingMethod.ELMO:
self.embedding = ElmoEmbedding()
elif method == EmbeddingMethod.ELMO_OFFLINE:
self.embedding = ElmoEmbeddingOffline(elmo_file)
super(WordEmbeddingRelationExtraction, self).__init__()