rasa nlu ner_crf not extracting any entities - tensorflow

Question

I have successfully got my code to detect the correct intent but no entities appear even though I provided some entities in my training data.
data.json:
{ “common_examples”: [
{ “text”:“Hello”,
“intent”:“greeting”,
“entities”:[] },
{ “text”:“Hi”,
“intent”:“greeting”,
“entities”:[] },
{ “text”:“I want a recipe for my lunch”,
“intent”:“get_recipe”,
“entities”:[
{ “start”:22,
“end”: 28,
“value”: “lunch”,
“entity”: “mealtime” }
]
},
{ “text”:“Can you give me a recipe for dinner tonight?”,
“intent”:“get_recipe”,
“entities”:[
{ “start”:29,
“end”:35,
“value”: “dinner”,
“entity”: “mealtime” }
]
},
{ “text”:“I don’t know what to have for lunch”,
“intent”:“get_recipe”,
“entities”:[
{ “start”:31,
“end”: 35,
“value”: “lunch”,
“entity”: “mealtime” }
]
}
},
}
],
"regex_features": [],
"entity_synonyms":[]
}
}
This is just a snippet. I have created 15 examples in total for the get_recipe intent. I just need it to pick out the entity of ‘mealtime’ from the message put to the bot.
My config.yml is as follows:
language: “en”
pipeline:
-name: “nlp_spacy”
-name: “tokenizer_spacy”
-name: “intent_entity_featurizer_regex”
-name: “intent_featurizer_spacy”
-name: “ner_crf”
-name: “ner_synonyms”
-name: “intent_featurizer_count_vectors”
-name: “intent_classifier_tensorflow_embedding”
and this is the code I run to train the bot:
from rasa_nlu.training_data import load_data
from rasa_nlu.model import Trainer
from rasa_nlu import config
from rasa_nlu.model import Interpreter
def train_bot(data_json,config_file,model_dir):
training_data = load_data(data_json)
trainer = Trainer(config.load(config_file))
trainer.train(training_data)
model_directory=trainer.persist(model_dir,fixed_model_name=‘vegabot’)
This runs fine.
And the code I run to predict the intent:
def predict_intent(text):
interpreter = Interpreter.load(‘models/nlu/default/vegabot’)
print(interpreter.parse(text))
Which produces the result:
{‘intent’: {‘name’: ‘get_recipe’, ‘confidence’: 0.9701309204101562}, ‘entities’: [], ‘intent_ranking’: [{‘name’: ‘get_recipe’, ‘confidence’: 0.9701309204101562}, {‘name’: ‘greeting’, ‘confidence’: 0.03588612377643585}], ‘text’: ‘can you find me a recipe for dinner’}
As you can see the intent is correct but entities is blank [] and I can’t figure out why. I don't seem to be getting any errors. Everything runs okay apart from this!
I also ran an evaluation and got:
- intent examples: 12 (2 distinct intents)
- Found intents: ‘greeting’, ‘get_recipe’
- entity examples: 10 (1 distinct entities)
- found entities: ‘mealtime’ which all looks fine.
So obviously it knows to look out for the mealtime entity but why isn't it picking it up from my test messages?
e.g. I need a recipe for lunch, Can you give me a dinner time recipe? etc
I’m using RASA NLU version 0.14.
Any help would be greatly appreciated. Thank you.

Answer 1

The machine learning models in Rasa need a bit of data to train. As correctly suggested in the comments, you have to give the conditional random field a couple of examples, so that it is actually able to generalize. Also make sure to vary the sentences around it, otherwise crf will not generalize to other contexts.