Pythonパッケージazure-ai-textanalyticsの6.Xを使う

Pythonパッケージazure-ai-textanalyticsの6.0.0b2を使ってみました。
以下には5.Xの書き方があるのですが、6.X の書き方が見つかりにくかったので、記録に残しておきます。

シナリオ

以下の３つのAPIを使っています。5.Xだと「個人を特定できる情報の検出」でカテゴリ(名前や組織等)の指定ができなかったので、プレリリースながら6.Xを使いました。

• キーフレーズの抽出
• 感情分析
• 個人を特定できる情報の検出

Python

UBuntu24.0.4でPython3.14.2を使っています。
前提としてMicrosoft Foundryリソースを作成しています。
5.xと比べると、すべてのAPIがanalyze_text関数に集約されていてbodyに渡す内容でコントロールしているのがわかります。

from azure.ai.textanalytics import TextAnalysisClient
from azure.core.credentials import AzureKeyCredential
from azure.ai.textanalytics.models import (
    AnalyzeTextKeyPhraseResult,
    ConfidenceScoreThreshold,
    EntityMaskPolicyType,
    KeyPhraseActionContent,
    PiiActionContent,
    MultiLanguageTextInput,
    MultiLanguageInput,
    TextPiiEntitiesRecognitionInput,
    TextKeyPhraseExtractionInput,
    AnalyzeTextPiiResult,
    TextSentimentAnalysisInput,
    AnalyzeTextSentimentResult,    
)

language_key = "<key>"
language_endpoint = "https://<resource>.cognitiveservices.azure.com/"

# Authenticate the client using your key and endpoint 
def authenticate_client():
    ta_credential = AzureKeyCredential(language_key)
    text_analytics_client = TextAnalysisClient(
            endpoint=language_endpoint, 
            credential=ta_credential)
    return text_analytics_client
 
client = authenticate_client()

def pre_analyze_text(kind: str, text: str)-> tuple:
    """
    Prepares and calls the unified text analysis function.
    
    Args:
        kind: Type of analysis - "sentiment", "key_phrase", or "pii"
        text: Text to analyze
    """
    # Common input structure
    common_input = MultiLanguageTextInput(
        multi_language_inputs=[MultiLanguageInput(id="1", text=text, language="ja")]
    )
    
    # Build body based on kind
    if kind == "sentiment":
        body = TextSentimentAnalysisInput(text_input=common_input)
        result_type = AnalyzeTextSentimentResult
    elif kind == "key_phrase":
        body = TextKeyPhraseExtractionInput(
            text_input=common_input,
            action_content=KeyPhraseActionContent(model_version="latest")
        )
        result_type = AnalyzeTextKeyPhraseResult
    elif kind == "pii":
        body = TextPiiEntitiesRecognitionInput(
            text_input=common_input,
            action_content=PiiActionContent(
                pii_categories=["Organization", "Email", "URL", "PhoneNumber", "Person", "Address"],
                redaction_policies=[
                    EntityMaskPolicyType(policy_name="entityMask", is_default=True),
                ],
                confidence_score_threshold=ConfidenceScoreThreshold(default=0.5)
            )
        )
        result_type = AnalyzeTextPiiResult
    else:
        print(f"Unknown kind: {kind}. Use 'sentiment', 'key_phrase', or 'pii'.")
        raise ValueError("Invalid analysis kind.")
    return body, result_type


# Unified text analysis function
def analyze_text(client, kind: str, text: str):
    """
    Unified function for text analysis.
    
    Args:
        client: TextAnalysisClient instance
        kind: Type of analysis - "sentiment", "key_phrase", or "pii"
        text: Text to analyze
    """

    try:
        body, result_type = pre_analyze_text(kind, text)
        result = client.analyze_text(body=body)
        
        if not isinstance(result, result_type) or not result.results or not result.results.documents:
            print("No documents in the response or unexpected result type.")
            return result
        
        # Output based on kind
        print(f"\nDocument ID: {result.results.documents[0].id}")
        
        if kind == "sentiment":
            return result.results.documents[0].sentiment.value
        elif kind == "key_phrase":
            if result.results.documents[0].key_phrases:
                return str(result.results.documents[0].key_phrases)
            else:
                return None
        elif kind == "pii":
            return result.results.documents[0].redacted_text        

    except Exception as e:
        print(f"Error executing analysis: {e}")
        return None

# Test the unified function
text_sample = "Microsoft 人事部の山田さんには080-1234-5678か[e-mail]で連絡して。"

print("=== Sentiment Analysis ===")
print(analyze_text(client, "sentiment", text_sample))

print("\n=== Key Phrase Extraction ===")
print(analyze_text(client, "key_phrase", text_sample))

print("\n=== PII Recognition ===")
print(analyze_text(client, "pii", text_sample))

以下は実行結果です。

=== Sentiment Analysis ===

Document ID: 1
neutral

=== Key Phrase Extraction ===

Document ID: 1
['Microsoft 人事部', '山田さん']

=== PII Recognition ===

Document ID: 1
[ORGANIZATION-1] 人事部の[PERSON-2]さんには[PHONENUMBER-3]か[EMAIL-4]で連絡して。