如何使用RASA NLU处理复合类型实体？以姓名实体场景为例

Absolutely, Rasa does support handling composite entities like the full name you mentioned (John James Doe broken into first name, middle name, and last name)! Here are a few practical approaches to implement this based on your needs:

1. Use Entity Roles (Built-in, Simplest Approach) #

Rasa's native entity roles let you tag sub-parts of a composite entity with specific roles, keeping them under a parent entity type. This is perfect for your name-splitting use case.

Example Training Data #

{
  "rasa_nlu_data": {
    "common_examples": [
      {
        "text": "My name is John James Doe",
        "intent": "Introduction",
        "entities": [
          {
            "start": 11,
            "end": 15,
            "value": "John",
            "entity": "Name",
            "role": "first_name"
          },
          {
            "start": 16,
            "end": 21,
            "value": "James",
            "entity": "Name",
            "role": "middle_name"
          },
          {
            "start": 22,
            "end": 25,
            "value": "Doe",
            "entity": "Name",
            "role": "last_name"
          }
        ]
      }
    ]
  }
}

How to Access the Data #

In your custom actions or dialogue management logic, you can retrieve specific parts using the role:

# Get the first name
first_name = next((e["value"] for e in tracker.latest_message["entities"] if e["entity"] == "Name" and e["role"] == "first_name"), None)

2. Build a Custom NER Component (For Automatic Splitting) #

If you don't want to manually tag every sub-part of the composite entity, you can create a custom component that splits a recognized parent entity (like the full Name entity from your original data) into its sub-components automatically.

Step 1: Create the Custom Component #

from rasa.nlu.components import Component
from rasa.shared.nlu.training_data.message import Message
from typing import List, Optional, Dict

class SplitFullNameComponent(Component):
    def __init__(self, component_config: Optional[Dict] = None):
        super().__init__(component_config)

    def process(self, messages: List[Message], **kwargs):
        for message in messages:
            entities = message.get("entities", [])
            updated_entities = []
            
            for entity in entities:
                if entity["entity"] == "Name":
                    name_parts = entity["value"].split()
                    name_length = len(name_parts)
                    
                    # Add first name
                    first_start = entity["start"]
                    first_end = first_start + len(name_parts[0])
                    updated_entities.append({
                        "start": first_start,
                        "end": first_end,
                        "value": name_parts[0],
                        "entity": "first_name",
                        "extractor": "custom_name_splitter"
                    })
                    
                    # Add middle names (if any)
                    if name_length > 2:
                        current_pos = first_end + 1
                        for part in name_parts[1:-1]:
                            mid_start = current_pos
                            mid_end = mid_start + len(part)
                            updated_entities.append({
                                "start": mid_start,
                                "end": mid_end,
                                "value": part,
                                "entity": "middle_name",
                                "extractor": "custom_name_splitter"
                            })
                            current_pos = mid_end + 1
                    
                    # Add last name
                    last_start = entity["end"] - len(name_parts[-1])
                    updated_entities.append({
                        "start": last_start,
                        "end": entity["end"],
                        "value": name_parts[-1],
                        "entity": "last_name",
                        "extractor": "custom_name_splitter"
                    })
                else:
                    updated_entities.append(entity)
            
            message.set("entities", updated_entities, add_to_output=True)

Step 2: Register the Component in config.yml #

Add your custom component to the NLU pipeline after the standard NER components:

language: en
pipeline:
  - name: WhitespaceTokenizer
  - name: LexicalSyntacticFeaturizer
  - name: CountVectorsFeaturizer
  - name: DIETClassifier
    epochs: 100
  - name: EntitySynonymMapper
  - name: "path.to.your.SplitFullNameComponent"  # Replace with your actual file path

3. Use Entity Groups (For Associating Sub-Entities) #

If you need to link multiple sub-entities to indicate they belong to the same composite entity (e.g., distinguishing between two people's names in one message), use entity groups.

Example Training Data #

{
  "rasa_nlu_data": {
    "common_examples": [
      {
        "text": "My name is John James Doe",
        "intent": "Introduction",
        "entities": [
          {
            "start": 11,
            "end": 15,
            "value": "John",
            "entity": "first_name",
            "group": "full_name_1"
          },
          {
            "start": 16,
            "end": 21,
            "value": "James",
            "entity": "middle_name",
            "group": "full_name_1"
          },
          {
            "start": 22,
            "end": 25,
            "value": "Doe",
            "entity": "last_name",
            "group": "full_name_1"
          }
        ]
      }
    ]
  }
}

This lets you group all sub-entities under a single identifier, making it easy to associate them in your logic.

内容的提问来源于stack exchange，提问作者Kunal Mukherjee