def test_count_vector_featurizer(sentence, expected):
from rasa.nlu.featurizers.count_vectors_featurizer import CountVectorsFeaturizer
ftr = CountVectorsFeaturizer({"token_pattern": r"(?u)\b\w+\b"})
train_message = Message(sentence)
# this is needed for a valid training example
train_message.set("intent", "bla")
data = TrainingData([train_message])
ftr.train(data)
test_message = Message(sentence)
ftr.process(test_message)
assert np.all(test_message.get("text_features") == expected)
def tokenize(self, message: Message, attribute: Text) -> List[Token]:
import MicroTokenizer
text = message.get(attribute)
tokenized = MicroTokenizer.cut(text)
tokens = []
offset = 0
for word in tokenized:
tokens.append(Token(word, offset))
offset += len(word)
return tokens
def test_do_not_overwrite_any_entities():
message = Message("Max lives in Berlin.")
message.set(ENTITIES, [{
"entity": "person",
"value": "Max",
"start": 0,
"end": 3
}])
training_data = TrainingData()
training_data.training_examples = [
Message("Hi Max!",
data={"entities": [{
"entity": "person",
"value": "Max"
}]}),
Message(
"I live in Berlin",
data={"entities": [{
"entity": "city",
"value": "Berlin"
}]},
),
]
training_data.lookup_tables = [{
"name":
"city",
"elements": ["London", "Berlin", "Amsterdam"]
}]
entity_extractor = RegexEntityExtractor()
entity_extractor.train(training_data)
entity_extractor.process(message)
entities = message.get(ENTITIES)
assert entities == [
{
"entity": "person",
"value": "Max",
"start": 0,
"end": 3
},
{
"entity": "city",
"value": "Berlin",
"start": 13,
"end": 19,
"extractor": "RegexEntityExtractor",
},
]
def process(self, message: Message, **kwargs: Any):
"""Process an incoming message.
This is the components chance to process an incoming
message. The component can rely on
any context attribute to be present, that gets created
by a call to :meth:`components.Component.pipeline_init`
of ANY component and
on any context attributes created by a call to
:meth:`components.Component.process`
of components previous to this one."""
# TODO 分词, 如果利用其它分词组件, 需要进一步调整
if not message.get("tokens", default=None):
self.extract_tokens(message)
# 词性标注
self.extract_poses(message)
# 句法依存
self.extract_parses(message)
# 抽取实体<序列标注+实体提取>
self.extract_entities(message)
# 抽取代词
self.extract_pronouns(message)
else:
# rasa tokenizers
tokens = message.get("tokens")
message.set("tokenizers", tokens)
# List tokens
tokens = [tokenizer_extract(token) for token in tokens]
message.set("tokens", tokens)
self.extract_poses(message)
# 句法依存
self.extract_parses(message)
# 抽取实体<序列标注+实体提取>
# 语义分割 ->
self.entity_segment(message)
# 属性分析 ->
self.link_analyze(message)
def test_lookup_tables(sentence, expected, labeled_tokens, spacy_nlp):
from rasa.nlu.featurizers.sparse_featurizer.regex_featurizer import RegexFeaturizer
lookups = [
{
"name":
"drinks",
"elements":
["mojito", "lemonade", "sweet berry wine", "tea", "club?mate"],
},
{
"name": "plates",
"elements": "data/test/lookup_tables/plates.txt"
},
]
ftr = RegexFeaturizer()
ftr.add_lookup_tables(lookups)
# adds tokens to the message
component_config = {"name": "SpacyTokenizer"}
tokenizer = SpacyTokenizer(component_config)
message = Message(sentence)
message.set("text_spacy_doc", spacy_nlp(sentence))
tokenizer.process(message)
sequence_features, sentence_features = ftr._features_for_patterns(
message, TEXT)
assert np.allclose(sequence_features.toarray(), expected[:-1], atol=1e-10)
assert np.allclose(sentence_features.toarray(), expected[-1], atol=1e-10)
# the tokenizer should have added tokens
assert len(message.get(TOKENS_NAMES[TEXT], [])) > 0
# the number of regex matches on each token should match
for i, token in enumerate(message.get(TOKENS_NAMES[TEXT])):
token_matches = token.get("pattern").values()
num_matches = sum(token_matches)
assert num_matches == labeled_tokens.count(i)
def tokens_without_cls(
message: Message, attribute: Text = TEXT
) -> Optional[List[Token]]:
"""Return tokens of given message without __CLS__ token.
All tokenizers add a __CLS__ token to the end of the list of tokens for
text and responses. The token captures the sentence features.
Args:
message: The message.
attribute: Return tokens of provided attribute.
Returns:
Tokens without CLS token.
"""
# return all tokens up to __CLS__ token for text and responses
if attribute in DENSE_FEATURIZABLE_ATTRIBUTES:
tokens = message.get(TOKENS_NAMES[attribute])
if tokens is not None:
return tokens[:POSITION_OF_CLS_TOKEN]
return None
# we don't add the __CLS__ token for intents, return all tokens
return message.get(TOKENS_NAMES[attribute])
def process(self, message: Message, **kwargs: Any) -> None:
"""Process an incoming message"""
entities = list(message.get('entities'))
# Get file path of lookup table in json format
cur_path = os.path.dirname(__file__)
if os.name == 'nt':
partial_lookup_file_path = '..\\data\\lookup_table.json'
else:
partial_lookup_file_path = '../data/lookup_table.json'
lookup_file_path = os.path.join(cur_path, partial_lookup_file_path)
with open(lookup_file_path, 'r') as file:
lookup_data = json.load(file)
tokens = message.get('tokens')
for token in tokens:
similarity_score = self.get_fuzzy_similarity(
token.text, lookup_data, self.threshold)
if similarity_score is not None:
print("'" + token.text + "'" + " matches with " +
str(similarity_score[0]) + "[" +
similarity_score[2] + "]" + " with a score of: " +
str(similarity_score[1]))
for i, item in enumerate(entities):
# if entity already exist, update it (because diet classifier is higher in hierarchy)
if item['entity'] == similarity_score[2]:
item.update({"value": similarity_score[0]})
entities.append({
"start": token.start,
"end": token.end,
"value": similarity_score[0],
"confidence": similarity_score[1],
"entity": similarity_score[2]
})
file.close()
message.set("entities", entities, add_to_output=True)
def test_mitie_featurizer_train(mitie_feature_extractor):
featurizer = MitieFeaturizer.create({}, RasaNLUModelConfig())
sentence = "Hey how are you today"
message = Message(sentence)
message.set(RESPONSE, sentence)
message.set(INTENT, "intent")
MitieTokenizer().train(TrainingData([message]))
featurizer.train(
TrainingData([message]),
RasaNLUModelConfig(),
**{"mitie_feature_extractor": mitie_feature_extractor},
)
expected = np.array(
[0.00000000e00, -5.12735510e00, 4.39929873e-01, -5.60760403e00, -8.26445103e00]
)
expected_cls = np.array([0.0, -4.4551446, 0.26073121, -1.46632245, -1.84205751])
vecs = message.get(DENSE_FEATURE_NAMES[TEXT])
assert len(message.get(TOKENS_NAMES[TEXT])) == len(vecs)
assert np.allclose(vecs[0][:5], expected, atol=1e-5)
assert np.allclose(vecs[-1][:5], expected_cls, atol=1e-5)
vecs = message.get(DENSE_FEATURE_NAMES[RESPONSE])
assert len(message.get(TOKENS_NAMES[RESPONSE])) == len(vecs)
assert np.allclose(vecs[0][:5], expected, atol=1e-5)
assert np.allclose(vecs[-1][:5], expected_cls, atol=1e-5)
vecs = message.get(DENSE_FEATURE_NAMES[INTENT])
assert vecs is None
def test_count_vector_featurizer_response_attribute_featurization(
sentence, intent, response, intent_features, response_features):
ftr = CountVectorsFeaturizer({"token_pattern": r"(?u)\b\w+\b"})
tk = WhitespaceTokenizer()
train_message = Message(sentence)
# this is needed for a valid training example
train_message.set(INTENT_ATTRIBUTE, intent)
train_message.set(RESPONSE_ATTRIBUTE, response)
# add a second example that has some response, so that the vocabulary for
# response exists
second_message = Message("hello")
second_message.set(RESPONSE_ATTRIBUTE, "hi")
second_message.set(INTENT_ATTRIBUTE, "greet")
data = TrainingData([train_message, second_message])
tk.train(data)
ftr.train(data)
if intent_features:
assert (train_message.get(
SPARSE_FEATURE_NAMES[INTENT_ATTRIBUTE]).toarray()[0] ==
intent_features)
else:
assert train_message.get(
SPARSE_FEATURE_NAMES[INTENT_ATTRIBUTE]) is None
if response_features:
assert (train_message.get(
SPARSE_FEATURE_NAMES[RESPONSE_ATTRIBUTE]).toarray()[0] ==
response_features)
else:
assert train_message.get(
SPARSE_FEATURE_NAMES[RESPONSE_ATTRIBUTE]) is None
def _from_text_to_crf(self,
message: Message,
entities: List[Text] = None) -> List[CRFToken]:
"""Takes a sentence and switches it to crfsuite format."""
crf_format = []
if self.pos_features:
tokens = message.get(SPACY_DOCS[TEXT_ATTRIBUTE])
else:
tokens = message.get(TOKENS_NAMES[TEXT_ATTRIBUTE])
text_dense_features = self.__get_dense_features(message)
for i, token in enumerate(tokens):
pattern = self.__pattern_of_token(message, i)
entity = entities[i] if entities else "N/A"
tag = self.__tag_of_token(token) if self.pos_features else None
dense_features = (text_dense_features[i]
if text_dense_features is not None else [])
crf_format.append(
CRFToken(token.text, tag, entity, pattern, dense_features))
return crf_format
def process(self, message: Message, **kwargs: Any) -> None:
mitie_feature_extractor = kwargs.get("mitie_feature_extractor")
if not mitie_feature_extractor:
raise Exception("Failed to train 'MitieFeaturizer'. "
"Missing a proper MITIE feature extractor.")
ents = self.extract_entities(message.text,
self._tokens_without_cls(message),
mitie_feature_extractor)
extracted = self.add_extractor_name(ents)
extracted = self.clean_up_entities(message, extracted)
message.set(ENTITIES,
message.get(ENTITIES, []) + extracted,
add_to_output=True)
def _get_processed_message_tokens_by_attribute(
self, message: Message, attribute: Text = TEXT
) -> List[Text]:
"""Get processed text of attribute of a message"""
if message.get(attribute) is None:
# return empty list since sklearn countvectorizer does not like None
# object while training and predicting
return []
tokens = self._get_message_tokens_by_attribute(message, attribute)
tokens = self._process_tokens(tokens, attribute)
tokens = self._replace_with_oov_token(tokens, attribute)
return tokens
def test_custom_intent_symbol(text, expected_tokens):
component_config = {
"intent_tokenization_flag": True,
"intent_split_symbol": "+"
}
tk = MitieTokenizer(component_config)
message = Message(text)
message.set(INTENT, text)
tk.train(TrainingData([message]))
assert [t.text
for t in message.get(TOKENS_NAMES[INTENT])] == expected_tokens
def test_spacy_ner_featurizer_config(spacy_nlp):
from rasa.nlu.featurizers.spacy_featurizer import SpacyFeaturizer
sentence = "hi there friend"
doc = spacy_nlp(sentence)
spacy_config = {"ner_feature_vectors": False}
ftr = SpacyFeaturizer.create(spacy_config, RasaNLUModelConfig())
greet = {"intent": "greet", "text_features": [0.5]}
message = Message(sentence, greet)
message.set("spacy_doc", doc)
ftr._set_spacy_features(message)
ftr._set_spacy_ner_features(message)
vecs = np.array(message.get("ner_features"))
assert vecs.shape[0] == len(doc)
assert vecs.shape[1] == 0
def _from_text_to_crf(
self,
message: Message,
entities: List[Text] = None
) -> List[Tuple[Optional[Text], Optional[Text], Text, Dict[Text, Any],
Optional[Dict[Text, Any]], ]]:
"""Takes a sentence and switches it to crfsuite format."""
crf_format = []
if self.pos_features:
tokens = message.get("spacy_doc")
else:
tokens = message.get("tokens")
ner_features = (self.__additional_ner_features(message)
if self.use_ner_features else None)
for i, token in enumerate(tokens):
pattern = self.__pattern_of_token(message, i)
entity = entities[i] if entities else "N/A"
tag = self.__tag_of_token(token) if self.pos_features else None
custom_ner_features = ner_features[
i] if self.use_ner_features else None
crf_format.append(
(token.text, tag, entity, pattern, custom_ner_features))
return crf_format
def test_spacy_ner_featurizer(sentence, expected, spacy_nlp):
from rasa.nlu.featurizers.spacy_featurizer import SpacyFeaturizer
doc = spacy_nlp(sentence)
token_vectors = [t.vector for t in doc]
spacy_config = {"ner_feature_vectors": True}
ftr = SpacyFeaturizer.create(spacy_config, RasaNLUModelConfig())
greet = {"intent": "greet", "text_features": [0.5]}
message = Message(sentence, greet)
message.set("spacy_doc", doc)
ftr._set_spacy_features(message)
ftr._set_spacy_ner_features(message)
vecs = message.get("ner_features")[0][:5]
assert np.allclose(token_vectors[0][:5], vecs, atol=1e-4)
assert np.allclose(vecs, expected, atol=1e-4)
def _set_spacy_ner_features(self, message: Message):
"""If we want to use spacy as an NER featurizer, set token vectors"""
doc = message.get(MESSAGE_SPACY_FEATURES_NAMES[MESSAGE_TEXT_ATTRIBUTE])
if self.ner_feature_vectors:
ner_features = np.array([t.vector for t in doc])
else:
ner_features = np.array([[] for t in doc])
combined_features = self._combine_with_existing_features(
message,
ner_features,
MESSAGE_VECTOR_FEATURE_NAMES[MESSAGE_ENTITIES_ATTRIBUTE],
)
message.set(
MESSAGE_VECTOR_FEATURE_NAMES[MESSAGE_ENTITIES_ATTRIBUTE], combined_features
)
def test_convert_featurizer_train(component_builder):
tokenizer = component_builder.create_component_from_class(ConveRTTokenizer)
featurizer = component_builder.create_component_from_class(ConveRTFeaturizer)
sentence = "Hey how are you today ?"
message = Message(sentence)
message.set(RESPONSE, sentence)
tokens = tokenizer.tokenize(message, attribute=TEXT)
tokens = tokenizer.add_cls_token(tokens, attribute=TEXT)
message.set(TOKENS_NAMES[TEXT], tokens)
message.set(TOKENS_NAMES[RESPONSE], tokens)
featurizer.train(
TrainingData([message]), RasaNLUModelConfig(), tf_hub_module=tokenizer.module
)
expected = np.array([2.2636216, -0.26475656, -1.1358104, -0.49751878, -1.3946456])
expected_cls = np.array(
[1.0251294, -0.04053932, -0.7018805, -0.82054937, -0.75054353]
)
vecs = message.get(DENSE_FEATURE_NAMES[TEXT])
assert len(tokens) == len(vecs)
assert np.allclose(vecs[0][:5], expected, atol=1e-5)
assert np.allclose(vecs[-1][:5], expected_cls, atol=1e-5)
vecs = message.get(DENSE_FEATURE_NAMES[RESPONSE])
assert len(tokens) == len(vecs)
assert np.allclose(vecs[0][:5], expected, atol=1e-5)
assert np.allclose(vecs[-1][:5], expected_cls, atol=1e-5)
vecs = message.get(DENSE_FEATURE_NAMES[INTENT])
assert vecs is None
def test_count_vector_featurizer_oov_token(sentence, expected):
ftr = CountVectorsFeaturizer({
"token_pattern": r"(?u)\b\w+\b",
"OOV_token": "__oov__"
})
train_message = Message(sentence)
WhitespaceTokenizer().process(train_message)
data = TrainingData([train_message])
ftr.train(data)
test_message = Message(sentence)
ftr.process(test_message)
assert np.all(
test_message.get(SPARSE_FEATURE_NAMES[TEXT]).toarray()[0] == expected)
def _create_sparse_features(self, message: Message) -> None:
"""Convert incoming messages into sparse features using the configured
features."""
import scipy.sparse
# [:-1] to remove CLS token
tokens = message.get(TOKENS_NAMES[TEXT])[:-1]
sentence_features = self._tokens_to_features(tokens)
one_hot_feature_vector = self._features_to_one_hot(sentence_features)
sparse_features = scipy.sparse.coo_matrix(one_hot_feature_vector)
sparse_features = self._combine_with_existing_sparse_features(
message, sparse_features, feature_name=SPARSE_FEATURE_NAMES[TEXT])
message.set(SPARSE_FEATURE_NAMES[TEXT], sparse_features)
def test_lm_tokenizer_number_of_sub_tokens(text, expected_number_of_sub_tokens):
transformers_config = {"model_name": "bert"} # Test for one should be enough
transformers_nlp = HFTransformersNLP(transformers_config)
lm_tokenizer = LanguageModelTokenizer()
message = Message(text)
td = TrainingData([message])
transformers_nlp.train(td)
lm_tokenizer.train(td)
assert [
t.get(NUMBER_OF_SUB_TOKENS) for t in message.get(TOKENS_NAMES[TEXT])[:-1]
] == expected_number_of_sub_tokens
def _split_intent(self,
message: Message,
attribute: Text = INTENT) -> List[Token]:
text = message.get(attribute)
# for INTENT_RESPONSE_KEY attribute,
# first split by RESPONSE_IDENTIFIER_DELIMITER
if attribute == INTENT_RESPONSE_KEY:
intent, response_key = text.split(RESPONSE_IDENTIFIER_DELIMITER)
words = self._tokenize_on_split_symbol(
intent) + self._tokenize_on_split_symbol(response_key)
else:
words = self._tokenize_on_split_symbol(text)
return self._convert_words_to_tokens(words, text)
def set_fasttext_features(self, message: Message, attribute: Text = TEXT) -> None:
tokens = message.get(TOKENS_NAMES[attribute])
if not tokens:
return None
text_vector = self.model.get_word_vector(message.text)
word_vectors = [
self.model.get_word_vector(t.text)
for t in train_utils.tokens_without_cls(message, attribute)
]
X = np.array(word_vectors + [text_vector]) # remember, we need one for __CLS__
features = self._combine_with_existing_dense_features(
message, additional_features=X, feature_name=DENSE_FEATURE_NAMES[attribute]
)
message.set(DENSE_FEATURE_NAMES[attribute], features)
def process(self, message: Message, **kwargs: Any) -> None:
import tensorflow as tf
import numpy as np
real_result_dir = os.path.join(self.model_dir, self.result_dir)
# print(real_result_dir)
if self.predict_fn is None:
self.predict_fn = tf.keras.experimental.load_from_saved_model(
real_result_dir)
real_lookup_table_file = os.path.join(real_result_dir,
self.lookup_table_file)
# print(real_lookup_table_file)
if self.lookup_table is None:
with open(real_lookup_table_file, 'rt') as fd:
self.lookup_table = json.load(fd)
text_feature = message.get("text_features")
np_feature = np.array([text_feature])
predict_np_int = self.predict_fn.predict(np_feature)
intent_score = []
for intent_id, score in enumerate(predict_np_int[0]):
# convert np.float32 to vanilla float,
# if not it will cause json_dumps of ujson raise exception OverflowError: Maximum recursion level reached
# see https://github.com/esnme/ultrajson/issues/221
float_score = float(score)
intent_score.append((float_score, intent_id))
reversed_lookup_table = {
index: value
for value, index in self.lookup_table.items()
}
intent_str_score = [(k, reversed_lookup_table[v])
for k, v in intent_score]
sorted_intent_str_score = sorted(intent_str_score,
key=lambda x: x[0],
reverse=True)
# print(sorted_intent_str_score)
self._set_intent_output(message, sorted_intent_str_score)
def test_lm_tokenizer_custom_intent_symbol(text, expected_tokens):
component_config = {"intent_tokenization_flag": True, "intent_split_symbol": "+"}
transformers_config = {"model_name": "bert"} # Test for one should be enough
transformers_nlp = HFTransformersNLP(transformers_config)
lm_tokenizer = LanguageModelTokenizer(component_config)
message = Message(text)
message.set(INTENT, text)
td = TrainingData([message])
transformers_nlp.train(td)
lm_tokenizer.train(td)
assert [t.text for t in message.get(TOKENS_NAMES[INTENT])] == expected_tokens
def _create_sparse_features(self, message: Message) -> None:
"""Convert incoming messages into sparse features using the configured
features."""
import scipy.sparse
# [:-1] to remove CLS token
tokens = message.get(TOKENS_NAMES[TEXT])[:-1]
sentence_features = self._tokens_to_features(tokens)
one_hot_feature_vector = self._features_to_one_hot(sentence_features)
sparse_features = scipy.sparse.coo_matrix(one_hot_feature_vector)
final_features = Features(
sparse_features, TEXT, self.component_config[FEATURIZER_CLASS_ALIAS]
)
message.add_features(final_features)
def test_count_vector_featurizer(sentence, expected):
from rasa.nlu.featurizers.count_vectors_featurizer import \
CountVectorsFeaturizer
ftr = CountVectorsFeaturizer({"min_ngram": 1,
"max_ngram": 2,
"analyzer": 'char'})
train_message = Message(sentence)
# this is needed for a valid training example
train_message.set("intent", "bla")
data = TrainingData([train_message])
ftr.train(data)
test_message = Message(sentence)
ftr.process(test_message)
assert np.all(test_message.get("text_features") == expected)
async def rebuild_original_text(example: Message) -> str:
"""
Rebuilds original training text in Markdown form.
"""
original_entities = example.get("entities")
original_text = example.text
if original_entities:
original_text = list(original_text)
for entity in sorted(original_entities,
key=lambda x: x.get("start"),
reverse=True):
start = entity["start"]
end = entity["end"]
value = entity["value"]
name = entity["entity"]
original_text[start:end] = f"[{value}]({name})"
original_text = "".join(original_text)
return original_text
def testing_tokenizer(text, cls, lang='en'):
from rasa.nlu.training_data import TrainingData, Message
defaults = {
# Flag to check whether to split intents
"intent_tokenization_flag": False,
# Symbol on which intent should be split
"intent_split_symbol": "_",
# text will be tokenized with case sensitive as default
"case_sensitive": True,
"lang": lang,
}
tok = cls(defaults)
example = Message(text, {"intent": "wish", "entities": []})
# tokenizer
tok.process(example, x='.')
for token in example.get("tokens"):
print(token.text, token.offset)
def test_spacy_featurizer_cls_vector(spacy_nlp):
featurizer = SpacyFeaturizer.create({}, RasaNLUModelConfig())
sentence = "Hey how are you today"
message = Message(sentence)
message.set(SPACY_DOCS[TEXT_ATTRIBUTE], spacy_nlp(sentence))
featurizer._set_spacy_features(message)
vecs = message.get(DENSE_FEATURE_NAMES[TEXT_ATTRIBUTE])
expected = np.array([-0.28451, 0.31007, -0.57039, -0.073056, -0.17322])
expected_cls = np.array(
[-0.196496, 0.3249364, -0.37408298, -0.10622784, 0.062756])
assert 6 == len(vecs)
assert np.allclose(vecs[0][:5], expected, atol=1e-5)
assert np.allclose(vecs[-1][:5], expected_cls, atol=1e-5)
