搭配 REST API 和 SynapseML 在網狀架構中使用預建文字分析（預覽）

發行項
11/15/2023

重要

這項功能處於預覽狀態。

文字分析是 Azure AI 服務，可讓您使用自然語言處理（NLP）功能來執行文字採礦和文字分析。

本教學課程示範如何在 Fabric 中使用文字分析搭配 RESTful API 來：

偵測句子或檔層級的情感標籤。
識別指定文字輸入的語言。
從文字擷取主要階段。
識別文字中的不同實體，並將其分類為預先定義的類別或類型。

必要條件

Rest API
SynapseML

# Get workload endpoints and access token

from synapse.ml.mlflow import get_mlflow_env_config
import json

mlflow_env_configs = get_mlflow_env_config()
access_token = access_token = mlflow_env_configs.driver_aad_token
prebuilt_AI_base_host = mlflow_env_configs.workload_endpoint + "cognitive/textanalytics/"
print("Workload endpoint for AI service: \n" + prebuilt_AI_base_host)

service_url = prebuilt_AI_base_host + "language/:analyze-text?api-version=2022-05-01"

# Make a RESful request to AI service

post_headers = {
    "Content-Type" : "application/json",
    "Authorization" : "Bearer {}".format(access_token)
}

def printresponse(response):
    print(f"HTTP {response.status_code}")
    if response.status_code == 200:
        try:
            result = response.json()
            print(json.dumps(result, indent=2, ensure_ascii=False))
        except:
            print(f"pasre error {response.content}")
    else:
        print(response.headers)
        print(f"error message: {response.content}")

import synapse.ml.core
from synapse.ml.cognitive.language import AnalyzeText
from pyspark.sql.functions import col

情感分析

Rest API
SynapseML

情感分析功能提供一種方式來偵測情感卷標（例如「負面」、「中性」和「正面」），以及句子和檔層級的信賴分數。此功能也會針對每份檔傳回 0 到 1 之間的信賴分數，以及針對正面、中性和負面情感的句子。如需啟用的語言清單，請參閱情感分析和意見採礦語言支援。

import requests
from pprint import pprint
import uuid

post_body = {
    "kind": "SentimentAnalysis",
    "parameters": {
        "modelVersion": "latest",
        "opinionMining": "True"
    },
    "analysisInput":{
        "documents":[
            {
                "id":"1",
                "language":"en",
                "text": "The food and service were unacceptable. The concierge was nice, however."
            }
        ]
    }
} 

post_headers["x-ms-workload-resource-moniker"] = str(uuid.uuid1())
response = requests.post(service_url, json=post_body, headers=post_headers)

# Output all information of the request process
printresponse(response)

輸出

    HTTP 200
    {
      "kind": "SentimentAnalysisResults",
      "results": {
        "documents": [
          {
            "id": "1",
            "sentiment": "mixed",
            "confidenceScores": {
              "positive": 0.43,
              "neutral": 0.04,
              "negative": 0.53
            },
            "sentences": [
              {
                "sentiment": "negative",
                "confidenceScores": {
                  "positive": 0.0,
                  "neutral": 0.01,
                  "negative": 0.99
                },
                "offset": 0,
                "length": 40,
                "text": "The food and service were unacceptable. ",
                "targets": [
                  {
                    "sentiment": "negative",
                    "confidenceScores": {
                      "positive": 0.01,
                      "negative": 0.99
                    },
                    "offset": 4,
                    "length": 4,
                    "text": "food",
                    "relations": [
                      {
                        "relationType": "assessment",
                        "ref": "#/documents/0/sentences/0/assessments/0"
                      }
                    ]
                  },
                  {
                    "sentiment": "negative",
                    "confidenceScores": {
                      "positive": 0.01,
                      "negative": 0.99
                    },
                    "offset": 13,
                    "length": 7,
                    "text": "service",
                    "relations": [
                      {
                        "relationType": "assessment",
                        "ref": "#/documents/0/sentences/0/assessments/0"
                      }
                    ]
                  }
                ],
                "assessments": [
                  {
                    "sentiment": "negative",
                    "confidenceScores": {
                      "positive": 0.01,
                      "negative": 0.99
                    },
                    "offset": 26,
                    "length": 12,
                    "text": "unacceptable",
                    "isNegated": false
                  }
                ]
              },
              {
                "sentiment": "positive",
                "confidenceScores": {
                  "positive": 0.86,
                  "neutral": 0.08,
                  "negative": 0.07
                },
                "offset": 40,
                "length": 32,
                "text": "The concierge was nice, however.",
                "targets": [
                  {
                    "sentiment": "positive",
                    "confidenceScores": {
                      "positive": 1.0,
                      "negative": 0.0
                    },
                    "offset": 44,
                    "length": 9,
                    "text": "concierge",
                    "relations": [
                      {
                        "relationType": "assessment",
                        "ref": "#/documents/0/sentences/1/assessments/0"
                      }
                    ]
                  }
                ],
                "assessments": [
                  {
                    "sentiment": "positive",
                    "confidenceScores": {
                      "positive": 1.0,
                      "negative": 0.0
                    },
                    "offset": 58,
                    "length": 4,
                    "text": "nice",
                    "isNegated": false
                  }
                ]
              }
            ],
            "warnings": []
          }
        ],
        "errors": [],
        "modelVersion": "2022-11-01"
      }
    }

情感分析功能提供一種方式來偵測情感卷標（例如「負面」、「中性」和「正面」），以及句子和檔層級的信賴分數。此功能也會針對其中每個檔與句子傳回介於 0 到 1 之間的信賴分數，以取得正面、中性和負面情感。如需啟用的語言清單，請參閱情感分析和意見採礦語言支援。

df = spark.createDataFrame([
    ("Great atmosphere. Close to plenty of restaurants, hotels, and transit! Staff are friendly and helpful.",),
    ("What a sad story!",)
], ["text"])

model = (AnalyzeText()
        .setTextCol("text")
        .setKind("SentimentAnalysis")
        .setOutputCol("response"))

result = model.transform(df)\
        .withColumn("documents", col("response.documents"))\
        .withColumn("sentiment", col("documents.sentiment"))

display(result.select("text", "sentiment"))

語言偵測器會評估每個檔的文字輸入，並傳回分數指出分析強度的語言標識元。此功能很適合用於收集未知語言任意文字的內容存放區。如需啟用的語言清單，請參閱支援的語言偵測語言。

post_body = {
    "kind": "LanguageDetection",
    "parameters": {
        "modelVersion": "latest"
    },
    "analysisInput":{
        "documents":[
            {
                "id":"1",
                "text": "This is a document written in English."
            }
        ]
    }
}

post_headers["x-ms-workload-resource-moniker"] = str(uuid.uuid1())
response = requests.post(service_url, json=post_body, headers=post_headers)

# Output all information of the request process
printresponse(response)

輸出

    HTTP 200
    {
      "kind": "LanguageDetectionResults",
      "results": {
        "documents": [
          {
            "id": "1",
            "detectedLanguage": {
              "name": "English",
              "iso6391Name": "en",
              "confidenceScore": 0.99
            },
            "warnings": []
          }
        ],
        "errors": [],
        "modelVersion": "2022-10-01"
      }
    }

df = spark.createDataFrame([
    (["Hello world"],),
    (["Bonjour tout le monde", "Hola mundo", "Tumhara naam kya hai?"],),
    (["你好"],),
    (["日本国（にほんこく、にっぽんこく、英"],)
], ["text"])

model = (AnalyzeText()
        .setTextCol("text")
        .setKind("LanguageDetection")
        .setOutputCol("response"))

result = model.transform(df)\
        .withColumn("documents", col("response.documents"))\
        .withColumn("detectedLanguage", col("documents.detectedLanguage.name"))

display(result.select("text", "detectedLanguage"))

關鍵片語擷取器

Rest API
SynapseML

關鍵片語擷取會評估非結構化文字，並傳回關鍵片語清單。如果您需要快速識別檔集合中的要點，這項功能就很有用。如需已啟用的語言清單，請參閱關鍵片語擷取支援的語言。

post_body = {
    "kind": "KeyPhraseExtraction",
    "parameters": {
        "modelVersion": "latest"
    },
    "analysisInput":{
        "documents":[
            {
                "id":"1",
                "language":"en",
                "text": "Dr. Smith has a very modern medical office, and she has great staff."
            }
        ]
    }
}

post_headers["x-ms-workload-resource-moniker"] = str(uuid.uuid1())
response = requests.post(service_url, json=post_body, headers=post_headers)

# Output all information of the request process
printresponse(response)

輸出

    HTTP 200
    {
      "kind": "KeyPhraseExtractionResults",
      "results": {
        "documents": [
          {
            "id": "1",
            "keyPhrases": [
              "modern medical office",
              "Dr. Smith",
              "great staff"
            ],
            "warnings": []
          }
        ],
        "errors": [],
        "modelVersion": "2022-10-01"
      }
    }

df = spark.createDataFrame([
    ("en", "Microsoft was founded by Bill Gates and Paul Allen."),
    ("en", "Text Analytics is one of the Azure Cognitive Services."),
    ("en", "My cat might need to see a veterinarian.")
], ["language", "text"])

model = (AnalyzeText()
        .setTextCol("text")
        .setKind("KeyPhraseExtraction")
        .setOutputCol("response"))

result = model.transform(df)\
        .withColumn("documents", col("response.documents"))\
        .withColumn("keyPhrases", col("documents.keyPhrases"))

display(result.select("text", "keyPhrases"))

具名實體辨識 (NER)

Rest API
SynapseML

具名實體辨識（NER）能夠識別文字中的不同實體，並將其分類為預先定義的類別或類型，例如：人員、位置、事件、產品和組織。如需已啟用的語言清單，請參閱 NER 語言支援。

post_body = {
    "kind": "EntityRecognition",
    "parameters": {
        "modelVersion": "latest"
    },
    "analysisInput":{
        "documents":[
            {
                "id":"1",
                "language": "en",
                "text": "I had a wonderful trip to Seattle last week."
            }
        ]
    }
}

post_headers["x-ms-workload-resource-moniker"] = str(uuid.uuid1())
response = requests.post(service_url, json=post_body, headers=post_headers)

# Output all information of the request process
printresponse(response)

輸出

    HTTP 200
    {
      "kind": "EntityRecognitionResults",
      "results": {
        "documents": [
          {
            "id": "1",
            "entities": [
              {
                "text": "trip",
                "category": "Event",
                "offset": 18,
                "length": 4,
                "confidenceScore": 0.74
              },
              {
                "text": "Seattle",
                "category": "Location",
                "subcategory": "GPE",
                "offset": 26,
                "length": 7,
                "confidenceScore": 1.0
              },
              {
                "text": "last week",
                "category": "DateTime",
                "subcategory": "DateRange",
                "offset": 34,
                "length": 9,
                "confidenceScore": 0.8
              }
            ],
            "warnings": []
          }
        ],
        "errors": [],
        "modelVersion": "2021-06-01"
      }
    }

df = spark.createDataFrame([
    ("en", "Microsoft was founded by Bill Gates and Paul Allen."),
    ("en", "Pike place market is my favorite Seattle attraction.")
], ["language", "text"])

model = (AnalyzeText()
        .setTextCol("text")
        .setKind("EntityRecognition")
        .setOutputCol("response"))

result = model.transform(df)\
        .withColumn("documents", col("response.documents"))\
        .withColumn("entityNames", col("documents.entities.text"))

display(result.select("text", "entityNames"))

實體連結

Rest API
SynapseML

本節中沒有 REST API 的步驟。

實體連結會識別並釐清在文字中找到之實體的身分識別。例如，在句子「我們上周前往西雅圖」中，會識別「Seattle」這個字，並連結到維琪百科的詳細資訊。如需已啟用的語言清單，請參閱實體連結的支持語言。

df = spark.createDataFrame([
    ("en", "Microsoft was founded by Bill Gates and Paul Allen."),
    ("en", "Pike place market is my favorite Seattle attraction.")
], ["language", "text"])

model = (AnalyzeText()
        .setTextCol("text")
        .setKind("EntityLinking")
        .setOutputCol("response"))

result = model.transform(df)\
        .withColumn("documents", col("response.documents"))\
        .withColumn("entityNames", col("documents.entities.name"))

display(result)

共用方式為

搭配 REST API 和 SynapseML 在網狀架構中使用預建文字分析（預覽）

必要條件

情感分析

輸出

語言偵測器

輸出

關鍵片語擷取器

輸出

具名實體辨識 (NER)

輸出

實體連結

意見反應

意見反應

其他資源

共用方式為

搭配 REST API 和 SynapseML 在網狀架構中使用預建 文字分析 （預覽）

必要條件

情感分析

輸出

語言偵測器

輸出

關鍵片語擷取器

輸出

具名實體辨識 (NER)

輸出

實體連結

相關內容

意見反應

意見反應

其他資源

搭配 REST API 和 SynapseML 在網狀架構中使用預建文字分析（預覽）