如何解决Rasa NLU 12类意图识别中的数据不平衡问题？

Got it, dealing with class imbalance in intent classification is super common—especially when you've got some intents with thousands of samples and others with just dozens. Let's break down the best ways to fix this for your Rasa setup, using your existing config as a starting point:

These target the root of the problem by balancing your training data:

• Oversample minority intents: For greetings, thanks, and other low-sample classes, manually add diverse, natural-looking examples. For instance, if you only have 20 greeting samples, expand with variations like "Hey there!", "Good afternoon!", "Hi, can you help me?" or use lightweight data augmentation (synonym swaps, rephrasing) to generate realistic new samples without overfitting.
• Carefully undersample majority intents: For high-volume intents like meeting setup, don’t just delete random samples. Instead, remove redundant or overly similar entries (e.g., duplicate "Schedule a meeting" phrases) while keeping a diverse set that covers all key scenarios (different meeting types, time frames, audiences).

Your current pipeline uses EmbeddingIntentClassifier—let’s adjust it to account for imbalance:

• Add class weights: The classifier supports a class_weight parameter that automatically assigns higher weights to underrepresented classes. Update your config like this:

  - name: "EmbeddingIntentClassifier"
    epochs: 100
    num_neg: 2
    class_weight: balanced
  

• Tune the num_neg parameter: Your current setting uses 2 negative samples per positive one. For imbalanced data, lower this to 1 to prevent the model from being dominated by majority-class samples.
• Swap to DIETClassifier (optional): If you want better robustness for both intent and entity tasks, replace EmbeddingIntentClassifier with DIETClassifier—it’s designed to handle imbalance better out of the box:

  - name: "DIETClassifier"
    epochs: 100
    class_weight: balanced
    entity_recognition: true
  

  (You can keep your CRFEntityExtractor and RegexEntityExtractor alongside this if needed.)

Make sure your core policies don’t overlook minority intents:

• Add RulePolicy for clear minority intents: For greetings, thanks, and other straightforward intents, define explicit rules to guarantee correct recognition. First, add RulePolicy to your policies list:

  policies:
    - name: MemoizationPolicy
    - name: RulePolicy
    - name: EmbeddingPolicy
      epochs: 20
    - name: FormPolicy
    - name: MappingPolicy
    - name: FallbackPolicy
      fallback_action_name: "action_default_fallback"
  

  Then create a rules.yml file with rules like:

  rules:
  - rule: Handle greeting intent
    steps:
    - intent: greet
    - action: utter_greet
  - rule: Handle thank you intent
    steps:
    - intent: thankyou
    - action: utter_thankyou
  

• Tweak FallbackPolicy thresholds: Lower the nlu_threshold (e.g., to 0.6) so the model triggers fallback when it’s unsure about an intent. This lets you collect unlabeled user inputs that might belong to your minority classes, which you can later annotate and add to your training data:

  - name: FallbackPolicy
    fallback_action_name: "action_default_fallback"
    nlu_threshold: 0.6
    core_threshold: 0.6
  

• Track per-class metrics: Don’t just rely on overall accuracy. Use rasa test nlu to generate a detailed report, and focus on precision, recall, and F1-score for your low-sample intents. This will show you exactly where the model is struggling.
• Active learning: Use the fallback interactions to collect new data for minority classes. Every time a user’s input triggers fallback, review it—if it’s a greeting or thank-you, label it and add it to your training set to gradually balance your data over time.

内容的提问来源于stack exchange，提问作者shaojie