如何将含意图与训练短语的CSV转为Rasa兼容nlu.md，或先建CSV再转换？

Alright, let's tackle these two Rasa NLU conversion challenges head-on. I’ve built similar workflows for clients before, so I’ll walk you through practical, reusable steps with examples you can adapt to your needs.

1. Convert a Specific CSV (Intent + Training Phrases) to Rasa's nlu.md #

First, let’s align on a common "specific format" CSV structure (since you didn’t provide exact specs, I’ll use a widely adopted setup). We’ll cover two common CSV variants: one phrase per row, and multiple phrases per intent row.

Step 1: Define Your CSV Structure #

Variant 1: Single Training Phrase Per Row

Save this as nlu_single_phrase.csv:

intent,training_phrase
greet,hi there
greet,hello!
greet,good morning
ask_weather,what's the weather today?
ask_weather,is it going to rain tomorrow?

Variant 2: Multiple Phrases Per Intent Row

Use a separator (like ;) to group phrases for the same intent. Save as nlu_multi_phrase.csv:

intent,training_phrases
greet,hi there;hello!;good morning
ask_weather,what's the weather today?;is it going to rain tomorrow?

Step 2: Python Script for Conversion #

This script handles both variants—no need to rewrite code for different CSV formats:

import csv

def csv_to_rasa_nlu(csv_path, output_md_path, multi_phrase_sep=None):
    """
    Convert intent-training phrase CSV to Rasa-compatible nlu.md
    :param csv_path: Path to your input CSV file
    :param output_md_path: Path to save the final nlu.md
    :param multi_phrase_sep: Separator for multiple phrases in one cell (e.g., ';')
    """
    intent_map = {}

    # Read and parse CSV
    with open(csv_path, 'r', encoding='utf-8') as csv_file:
        reader = csv.DictReader(csv_file)
        for row in reader:
            intent = row['intent'].strip()
            phrase_raw = row['training_phrases' if multi_phrase_sep else 'training_phrase'].strip()

            # Split phrases if using multi-phrase format
            phrases = [p.strip() for p in phrase_raw.split(multi_phrase_sep)] if multi_phrase_sep else [phrase_raw]
            phrases = [p for p in phrases if p]  # Remove empty strings

            # Build intent-phrase mapping
            if intent not in intent_map:
                intent_map[intent] = []
            intent_map[intent].extend(phrases)

    # Write to nlu.md
    with open(output_md_path, 'w', encoding='utf-8') as md_file:
        md_file.write("# NLU Training Data\n\n")
        for intent, phrases in intent_map.items():
            md_file.write(f"## intent:{intent}\n")
            for phrase in phrases:
                md_file.write(f"- {phrase}\n")
            md_file.write("\n")

# Usage for single-phrase CSV
csv_to_rasa_nlu('nlu_single_phrase.csv', 'nlu.md')

# Usage for multi-phrase CSV (uncomment below)
# csv_to_rasa_nlu('nlu_multi_phrase.csv', 'nlu.md', multi_phrase_sep=';')

Step 3: Verify the Output #

Your nlu.md will match Rasa’s required format exactly:

# NLU Training Data

## intent:greet
- hi there
- hello!
- good morning

## intent:ask_weather
- what's the weather today?
- is it going to rain tomorrow?

2. Design a Multi-Intent CSV & Convert to nlu.md #

Rasa supports multi-intents using the + separator (e.g., greet+ask_weather). Let’s design a CSV that handles both multi-intent labels and grouped training phrases.

Step 1: Multi-Intent CSV Structure #

We’ll stick to two columns, but allow intent to be a combined label and training_phrases to hold multiple relevant sentences. Save as nlu_multi_intent.csv:

intent,training_phrases
greet,hi there;hello!;good morning
ask_weather,what's the weather today?;is it going to rain tomorrow?
greet+ask_weather,hi, what's the weather today?;hello, is it going to rain tomorrow?;good morning, can you tell me the forecast?
order_pizza+order_drink,i want a pepperoni pizza and a coke;can i get a margherita pizza plus a sprite?

Step 2: Convert with the Same Script #

Reuse the script from part 1—just pass the multi-phrase separator:

csv_to_rasa_nlu('nlu_multi_intent.csv', 'nlu_multi_intent.md', multi_phrase_sep=';')

Step 3: Final nlu_multi_intent.md Output #

The file will correctly format multi-intents for Rasa:

# NLU Training Data

## intent:greet
- hi there
- hello!
- good morning

## intent:ask_weather
- what's the weather today?
- is it going to rain tomorrow?

## intent:greet+ask_weather
- hi, what's the weather today?
- hello, is it going to rain tomorrow?
- good morning, can you tell me the forecast?

## intent:order_pizza+order_drink
- i want a pepperoni pizza and a coke
- can i get a margherita pizza plus a sprite?

Quick Tips for Multi-Intent in Rasa #

• Ensure your pipeline in config.yml supports multi-intents (most standard pipelines do, but double-check if you’re using a custom setup)
• Reference combined intents directly in your stories/rules (e.g., intent:greet+ask_weather) to handle them in conversations

内容的提问来源于stack exchange，提问作者Lohith Arcot