How does RasaNLU perform entity extraction?
I started Demystifying Rasa NLU when I committed myself to #100DaysOfMLCode Challenge by Siraj Raval. For the first 10 days, I backtracked through the code base understanding what happens when we train the chatbot. By the 25th Day, I could crack what happens when we serve a trained model via RestAPIs.
With this overall understanding, I wanted to dig deeper and explore what happens during EntityExtraction. In this blog we will drill down through the entity extraction step in RasaNLU. This blog concentrates more on the structuring of data and building an ML model rather than understanding the statistical models that drive it from underneath.
RasaNLU & EntityExtraction
RasaNLU supports multiple methods to extract entities.
- NER CRF.
- NER Spacy
- NER Duckling
- NER MITIE
The use-case of each of these extractors is explained in RasaNLU’s documentation. Spacy and Duckling are commonly used for pre-trained entities like name, place, time,date etc.,
For custom entities like a product, cuisine, types of pizza we need to use NERCRF or MITIE. In this blog, we are going to stick with NERCRF since it is the default extractor provided by Rasa.
NERCRF — Named Entity Extraction using Conditional Random Field
Conditional Random Field(CRF) — a popular statistical model used in natural language processing, predicts sequences of labels for sequences of input samples.’
RasaNLU uses sklearncrfsuite to perform entity extraction, and you can use it outside the context of RasaNLU by following this tutorial.
Training Data
Training Sample{'intent': 'restaurant_search','entities': [{'start': 31, 'end': 37, 'value': 'centre', 'entity': 'location',{'start': 10, 'end': 17, 'value': 'mexican', 'entity': 'cuisine'}],'text': 'show me a mexican place in the centre'}
Preprocessing
To extract entities from the text, we need to Tokenize them. One thing I noticed here is that Unlike other NLP techniques, we don’t remove stop words for entity extraction. NERCRF is a sequential statistical model, i.e. every feature is dependent on features(tokens) before and after it and it is important to preserve the order of the word.
'tokens': [show, me, a, mexican, place, in, the, centre],
Create EntityExtractor Object
The CRFEntityExtractor object is created with default configuration settings provided by RasaNLU.
>>> crf = CRFEntityExtractor()# Default config>>> crf.component_config
{'BILOU_flag': True,'features': [
\['low', 'title', 'upper', 'pos', 'pos2'\],
\['bias', 'low', 'word3', 'word2', 'upper','title', 'digit','pos', 'pos2','pattern'\],
\['low', 'title', 'upper', 'pos', 'pos2'\]\],
'max\_iterations': 50,
'L1\_c': 1,
'L2\_c': 0.001
}
The features low, title, etc., are the features we will be extracting from the text. We will dig deeper into what each feature means later in this blog. L1_C and L2_C are parameters for the loss functions.
Filter Training data
RasaNLU supports more than one entity extractors.A bot developer can configure it using the extractor key in training example. If there is no extractor mentioned, NERCRF is taken as the default EntityExtractor. Hence, filter all training samples which does not have an extractor parameter or the one that specifies NERCRF.
def filter_trainable_entities(self, entity_examples):for ent in message.get("entities", []):extractor = ent.get("extractor")if not extractor or extractor == self.name:entities.append(ent)
Create Dataset
NERCRF uses a certain structure of training data. In this step we will transform RasaNLU's dataset into the one that is supported by NERCRF. This involves merging the entities with the training example associated with it.
Convert Entity
In the training data we have entities defined as dictionaries, In this step it is transformed to list of tuples with (start_index, end_index and entity). This will enable us to tag the tokens easily in the next step.
entities = [entity["start"], entity["end"], entity["entity"]for ent in example.get("entities", [])]
[(31, 37, ‘location’), (10, 17, ‘cuisine’)]
BILOU Tagging
BILOU means (Begin, Intermediate, Last, Other, Unigram) is a text tagging format that enables entity extraction. In this step for each token if it falls under the entity offset a entity tag is attached to it. If not tag O will be attached to it.
def _bilou_tags_from_offsets(tokens, entities, missing='O'):if start_token is not None and end_token is not None:** if start_token == end_token:bilou[start_token] = 'U-%s' % labelelse:bilou[start_token] = 'B-%s' % labelfor i in range(start_token + 1, end_token):bilou[i] = 'I-%s' % labelbilou[end_token] = 'L-%s' % label**
RasaNLU extracts both single and multi word entities. In case of single entity the Entity is tagged with U-Entity , A multi word entity has starting word tagged with B-Entityand Last word with Land all intermediate word with L CRFEntityExtractor.
tokens :: [show, me, a, mexican, place, in, the, centre]BILOU :: [O, O, O,'U-cuisine', O, O, O, 'U-location']
tokens :: [show, me, a, mexican, place, in, New, York, City]BILOU :: [O, O, O,'U-cuisine', O, O, O, 'B-location','I-location','L-location' ]
Text to CRF format
This is the final step of the dataset conversion where the BILOU tags and the tokens are merged to arrive at the following format.
[(token, POS_tag, entitiy, pattern)]
For each token in a training example a tuple of token its POSTag, BILOUEntity and RegexPattern in case of regex based entities.
>>> dataset[('show', None, 'O', None), ('me', None, 'O', None),('a', None, 'O', None), ('mexican', None, 'U-cuisine', None),('place', None, 'O', None), ('in', None, 'O', None),('the', None, 'O', None), ('centre', None, 'U-location', None)]
Create Training data
X_Train
NERCRF comes with a bunch of default configurations which we would use to generate features. As we seen before the default configuration looks something like this.
{'BILOU_flag': True,'features': [# Features to be extracted from token before['low', 'title', 'upper', 'pos', 'pos2'],# Features to be extracted from current token['bias', 'low', 'word3', 'word2', 'upper','title', 'digit','pos', 'pos2','pattern'],# Features to be extracted from token after ['low', 'title', 'upper', 'pos', 'pos2']],'max_iterations': 50,'L1_c': 1,'L2_c': 0.001}
If you look at the features list, we can see that there are 3 set of features. For each token in a training_example we would construct data for tokens before and after that particular token.
>>> X_train = [self._sentence_to_features(sent) for sent in dataset]>>> X_train[-1][...,{'-1:low': 'the', # The previous word'-1:title': False, # is the prev word a title'-1:upper': False, # is the prev word in uppercase'0:bias': 'bias','0:low': 'centre', # Current word'0:prefix5': 'centr','0:prefix2': 'ce','0:suffix5': 'entre','0:suffix3': 'tre','0:suffix2': 're','0:upper': False,'0:title': False, # is a title'0:digit': False, # is a number0:pattern': 'N/A','EOS': True} # End of sentence]
y_train
Y_train is the entity tags that we generated by BILOU tagging.
>>> y_train = [self._sentence_to_labels(sent) for sent in dataset]['O', 'O', 'O', 'U-cuisine', 'O', 'O', 'O', 'U-location']
Training
Now that we have the data ready all we have to do is create the CRF object and fit it using the training data. This step happens during the training phase of RasaNLU, and the prediction(next part) happens when we deploy the model.
crf = sklearn_crfsuite.CRF(algorithm='lbfgs',**component_config)crf.fit(X_train, y_train)
Prediction
During prediction, the text that the user enters is preprocessed and arrived at a format similar to the X_train. This preprocessed format is then fed into CRFClassifier to make predictions.
Preprocessing
training_sample = {'text': 'show me chines restaurants in the north'}# 1. Load into SpacyNLP{'spacy_doc': show me chines restaurants in the north,'text': 'show me chines restaurants in the north'}
# 2. Tokenize{'spacy_doc': "..."'tokens': [show, me, chines, restaurants, in, the, north],'text': "..."}
Next step is to convert the text into CRF format. Unlike training phase, the entity part of the list is filled with N/A since we won’t have any information on whether a particular token is an entity or not.
# 3. Convert Text to CRF format
>>> text_data = crf._from_text_to_crf(test_example)>>> text_data[('show', None, 'N/A', None),('me', None, 'N/A', None),('chines', None, 'N/A', None),('restaurants', None, 'N/A', None),('in', None, 'N/A', None),('the', None, 'N/A', None),('north', None, 'N/A', None)]
# 4. Convert CRF format to features>>> features = crf._sentence_to_features(text_data)>>> features
[{'BOS': True, '0:bias': 'bias', '0:low': 'show', '0:prefix5': 'show', '0:prefix2': 'sh', '0:suffix5': 'show', '0:suffix...
Predict
After prediction, we will get a list with the probability of a particular token being a BILOU-tag. This means we can ignore the tags with O and construct a dictionary with only valid entity tags.
# 1. Predict>>> ents = crf.ent_tagger.predict_marginals_single(features)>>> ents
[{'O': 0.9881565662477786, 'U-location': 0.0019499979138974536, 'U-cuisine': 0.002866549564514255, 'B-cusine': 0.00126939581550308, 'L-cusine': 0.002386854107682877, 'I-cusine': 0.0014001745853175994, 'U-cusine': 0.0019704617653066633}...]
# 2.Convert prediction result to readable format
>>> entities = crf._from_crf_to_json(test_example, ents))>>> entities
[{'start': 8, 'end': 14, 'value': 'chines', 'entity': 'cuisine', 'confidence': 0.848068806858813}, {'start': 34, 'end': 39, 'value': 'north', 'entity': 'location', 'confidence': 0.9334140282928288}]
With this, we have demystified one major component of any NLP problem — Entity extraction.
In the next part, We will digger deeper and understand the Intent Classification.
Was this post useful to you? Hold the clap button and give a shout out to me on twitter. ❤️