Manipulating message in parse function of model.py

Rasa Open Source
1

I want to ask if this is right how I am manipulate the incoming message from core before parsing with NLU. I do this within the parse function of model.py where I just use a simple stemmer. Now I get different intent for

problem with data

problems with data

Although both are get stemmed to same way! the word problems is stemmed same like problem.

I wonder why? Same sentences gets parsed afterwards in the components. Are there some caveats?

After parsing I cehck if intent is None from tensorflow and want to overwrite it. Although the tracker are seem to get correctly filled, I see some difference when uttering a None word vs entering \BLA ?

    def parse(self, text, time=None, only_output_properties=True):
        # type: (Text) -> Dict[Text, Any]
        """Parse the input text, classify it and return pipeline result.

        The pipeline result usually contains intent and entities."""
        stemmer= Stemmer()


        if not text:
            # Not all components are able to handle empty strings. So we need
            # to prevent that... This default return will not contain all
            # output attributes of all components, but in the end, no one
            # should pass an empty string in the first place.
            output = self.default_output_attributes()
            output["text"] = ""
            return output
        tokens = nltk.word_tokenize(text)
        tokens = [stemmer.stem(t) for t in tokens]
        text=" ".join(tokens)
        #logger.info("Parsed Message:************** '{}'".format(text))
        #print(text)
        message = Message(text, self.default_output_attributes(), time=time)

        for component in self.pipeline:
            component.process(message, **self.context)
        
       
        if message.data['intent']['name']==None:
           # print("YES")
            m=text.split()
            if len(m)==1:                
                message.data['intent']['name']='BLA'
                message.data['intent']['confidence']=0.99
      #  print(message.data['intent'])


        output = self.default_output_attributes()
        output.update(message.as_dict(
                only_output_properties=only_output_properties))
        return output

Any idea @akelad

:slight_smile:

2

Here is the log showing a real difference in tracker states. I input a text which is None intent and compare it versus /Bla directly to which I want to set a None word to like above.

vbn
DEBUG:rasa_core.tracker_store:Creating a new tracker for id 'default'.
DEBUG:rasa_core.tracker_store:Recreating tracker for id 'default'
vbn
DEBUG:rasa_core.processor:Received user message 'vbn' with intent '{'name': 'BLA', 'confidence': 0.99}' and entities '[]'
parsedata
{'intent': {'name': 'BLA', 'confidence': 0.99}, 'entities': [], 'intent_ranking': [], 'text': 'vbn'}
DEBUG:rasa_core.processor:Logged UserUtterance - tracker now has 2 events
DEBUG:rasa_core.processor:Current slot values:
...
NEXT ACTION
UserUttered(text: vbn, intent: {'name': 'BLA', 'confidence': 0.99}, entities: [])

CURRENT STAE
{'sender_id': 'default', 'slots': {...}, 'latest_message': {'intent': {'name': 'BLA', 'confidence': 0.99}, 'entities': [], 'intent_ranking': [], 'text': 'vbn'}, 'latest_event_time': 1541096797.4163308, 'paused': False, 'events': None}

trackers_as_states
[[None, {}, {'prev_action_listen': 1.0}]]

POLICY NAME
KerasPolicy
0.24645854532718658
POLICY NAME
FallbackPolicy
0.1
DEBUG:rasa_core.policies.ensemble:Predicted next action using policy_0_KerasPolicy
DEBUG:rasa_core.policies.ensemble:Predicted next action 'action_BLA' with prob 0.65.
ACTION NAME
action_BLA

compared to uttering with /BLA

/BLA
DEBUG:rasa_core.tracker_store:Recreating tracker for id 'default'
DEBUG:rasa_core.processor:Received user message '/BLA' with intent '{'name': 'BLA', 'confidence': 1.0}' and entities '[]'
parsedata
{'text': '/BLA', 'intent': {'name': 'BLA', 'confidence': 1.0}, 'intent_ranking': [{'name': 'BLA', 'confidence': 1.0}], 'entities': []}
DEBUG:rasa_core.processor:Logged UserUtterance - tracker now has 14 events
DEBUG:rasa_core.processor:Current slot values:
...
NEXT ACTION
UserUttered(text: /BLA, intent: {'name': 'BLA', 'confidence': 1.0}, entities: [])
CURRENT STAE
{'sender_id': 'default', 'slots': {...}, 'latest_message': {'text': '/BLA', 'intent': {'name': 'BLA', 'confidence': 1.0}, 'intent_ranking': [{'name': 'BLA', 'confidence': 1.0}], 'entities': []}, 'latest_event_time': 1541097292.5325198, 'paused': False, 'events': None}

trackers_as_states
[[None, {}, {'intent_BLA: 1.0, 'prev_action_listen': 1.0}]]

POLICY NAME
KerasPolicy
0.9567252397537231
POLICY NAME
FallbackPolicy
0.1
DEBUG:rasa_core.policies.ensemble:Predicted next action using policy_0_KerasPolicy
DEBUG:rasa_core.policies.ensemble:Predicted next action 'action_BLA' with prob 1.00.
ACTION NAME
action_BLA

So you see a difference in the tracker state:

trackers_as_states
[[None, {}, {'prev_action_listen': 1.0}]]

VS

trackers_as_states
[[None, {}, {'intent_BLA: 1.0, 'prev_action_listen': 1.0}]]

Manual overwriting the intent in model.py seems to miss something. The intent for a word with BLA intent coming from overwriting don’t yield the the bla intent inside the tracker as seen from the last log.

Therefore my Keras Policy does not work so well.

3

@akelad. I do it now with a custom component and the 2nd issue is solved. But I had to add a intent ranking not just name and confidence? So, adding a intent ranking in model.py would be suffiecient too.

Could you imagine what is the problem with my first issue at the beginning using the stemmer?

Thanks!