Rasa中意图未识别的处理方案及相关问题咨询

Hey there! Let's work through your Rasa weather bot issues step by step—first fixing the location intent recognition problem, then covering how to handle unrecognized intents and the role of stories.md.

First: Fixing Location Intent/Entity Recognition #

Your bot not picking up "Italy" even though it's in your data.json points to issues with Rasa NLU setup or training. Here's what to check:

1. Validate Your NLU Training Data (data.json) #

Make sure your inform intent samples are properly annotated with the location entity. For example:

{
  "text": "What's the weather like in Italy?",
  "intent": "inform",
  "entities": [{"start": 26, "end": 31, "value": "Italy", "entity": "location"}]
}

• Ensure you have multiple varied samples for the inform intent (e.g., "Weather in Rome?", "Tell me about London's forecast")—too few samples will make NLU less accurate.
• Double-check that entity start/end indices match the exact position of the location in the text (off-by-one errors break entity extraction).

2. Check Your NLU Pipeline Configuration #

For Rasa NLU 0.12.3, your nlu_config.yml needs a pipeline that includes intent classification and entity extraction. A working setup might look like:

language: "en"
pipeline:
- name: "nlp_spacy"
- name: "tokenizer_spacy"
- name: "intent_featurizer_spacy"
- name: "intent_classifier_sklearn"
- name: "entity_extractor_crf"

• Without an entity extractor like entity_extractor_crf, your bot won't pull the location value from user input even if the intent is recognized.

3. Re-Train & Verify NLU Model #

Run the NLU training command to ensure your model is up-to-date:

python -m rasa_nlu.train -c nlu_config.yml --data data.json -o models --fixed_model_name nlu --project current

After training, test the model directly to confirm it recognizes your input:

python -m rasa_nlu.server -c nlu_config.yml --model models/current/nlu

Then send a request (via the web interface or POST call) with "What's the weather in Italy?"—check if the response correctly identifies the inform intent and location entity.

4. Fix Domain File Issues #

Your domain file has inconsistencies that might break slot filling:

• Set action_factory: "default" (instead of null)—this tells Rasa Core how to handle actions.
• Add auto_fill: true to your location slot to ensure it automatically gets populated with the extracted entity:

  slots:
    location:
      initial_value: null
      type: rasa_core.slots.TextSlot
      auto_fill: true
  

• Remove the redundant action_names field—only the actions field is needed.

Handling Unrecognized Intents #

When your bot can't identify a user's intent, you'll want to gracefully fall back instead of failing silently. Here's how to set this up:

1. Add Fallback Intent to Domain #

Update your domain to include the nlu_fallback intent (triggered when NLU's intent confidence is below your threshold) and a fallback response template:

intents:
  - greet
  - goodbye
  - inform
  - nlu_fallback

templates:
  # ... your existing templates
  utter_please_rephrase:
    - text: "Sorry, I didn't catch that. Could you rephrase your question or ask about the weather in a specific location?"

2. Adjust NLU Confidence Threshold #

In your nlu_config.yml, set a threshold for the intent classifier—any intent with confidence below this will trigger nlu_fallback:

pipeline:
  # ... your existing pipeline steps
  - name: "intent_classifier_sklearn"
    threshold: 0.7  # Adjust based on your needs (0.5-0.8 is typical)

3. Add Fallback Story to stories.md #

Yes, you do need to add stories for unrecognized intents—Rasa Core uses stories to learn conversation flows. Add this story to handle fallbacks:

## Fallback path
* nlu_fallback
  - utter_please_rephrase

Bonus: Guard Against Empty API Calls #

Update your ActionWeather code to check if the location slot is empty before calling the weather API. This prevents empty API requests and prompts the user for a location if needed:

from rasa_core_sdk import Action
from rasa_core_sdk.events import SlotSet

class ActionWeather(Action):
    def name(self):
        return 'action_weather'

    def run(self, dispatcher, tracker, domain):
        location = tracker.get_slot('location')
        # Check if location is missing
        if not location:
            dispatcher.utter_template('utter_ask_location', tracker)
            return []
        # Proceed with API call
        # ... your weather API logic here ...
        dispatcher.utter_message(f"The weather in {location} is sunny with a high of 25°C!")
        return [SlotSet('location', location)]

内容的提问来源于stack exchange，提问作者Alice Antoine