基于Microsoft SemanticKernel和GPT4实现一个智能翻译服务

今年.NET Conf China 2023技术大会，我给大家分享了 .NET应用国际化-AIGC智能翻译+代码生成的议题

.NET Conf China 2023分享-.NET应用国际化-AIGC智能翻译+代码生成

今天将详细的代码实现和大家分享一下。

一、前提准备

1. 新建一个Console类的Project

2. 引用SK的Nuget包，SK的最新Nuget包

dotnet add package Microsoft.SemanticKernel --version 1.4.0

<ItemGroup>
    <PackageReference Include="Microsoft.SemanticKernel" Version="1.4.0" />
    <PackageReference Include="Newtonsoft.Json" Version="13.0.3" />
  </ItemGroup>

3. 在Azure OpenAI Service中创建一个GPT4的服务，这个可能大家没有账号，那就先看代码如何实现吧

部署好GPT4模型后，可以拿到以下三个重要的值

Azure OpenAI Deployment Name

Azure OpenAI Endpoint

Azure OpenAI Key

二、编写翻译使用的Prompt

 {{$input}}
请将上面的输入翻译为英文,不要返回任何解释说明,
请扮演一个美国电动汽车充电服务运营商(精通中文和英文),用户的输入数据是JSON格式,例如{"1":"充电站", "2":"充电桩"},
如果不是JSON格式,请返回无效的输入。
请使用以下专业术语进行翻译
 {
    "充电站":"Charging station",
    "电站":"Charging station",
    "场站":"Charging station",
    "充电桩":"Charging point",
    "充电终端":"Charging point",
    "终端":"Charging point",
    "电动汽车":"Electric Vehicle",
    "直流快充":"DC Fast Charger",
    "超级充电站":"Supercharger",
    "智能充电":"Smart Charging",
    "交流慢充":"AC Slow Charging"
}
翻译结果请以JSON格式返回,例如 {"1":"Charging station", "2":"Charging point"}

类似的还有葡萄牙下的翻译Prompt

{{$input}}
请将上面的输入翻译为葡萄牙语,不要返回任何解释说明,请扮演一个巴西的电动汽车充电服务运营商（精通葡萄牙语、中文和英文）
用户的输入数据是JSON格式,例如{"1":"充电站", "2":"充电桩"}, 如果不是JSON格式,请返回无效的输入
请使用以下专业术语进行翻译
 {
  "充电站": "Estação de carregamento",
  "电站": "Estação de carregamento",
  "场站": "Estação de carregamento",
  "充电桩": "Ponto de carregamento",
  "充电终端": "Ponto de carregamento",
  "终端": "Ponto de carregamento",
  "电动汽车": "Veículo Elétrico",
  "直流快充": "Carregador Rápido DC",
  "超级充电站": "Supercharger",
  "智能充电": "Carregamento Inteligente",
  "交流慢充": "Carregamento AC Lento"
}
请以JSON格式返回,例如 {"1":"Estação de carregamento", "2":"Ponto de carregamento"}

在项目工程下新建Plugins目录和TranslatePlugin子目录，同时新建Translator_en和Translator_pt等多个子目录

config.json文件下的内容如下：

{
    "schema": 1,
    "type": "completion",
    "description": "Translate.",
    "completion": {
         "max_tokens": 2000,
         "temperature": 0.5,
         "top_p": 0.0,
         "presence_penalty": 0.0,
         "frequency_penalty": 0.0
    },
    "input": {
         "parameters": [
              {
                   "name": "input",
                   "description": "The user's input.",
                   "defaultValue": ""
              }
         ]
    }
}

三、Translator翻译类，实现文本多语言翻译

这个类主要实现将用户输入的文本（系统处理为JSON格式），翻译为指定的语言

using System.Runtime.InteropServices;

using Microsoft.SemanticKernel;

using Newtonsoft.Json;

namespace LLM_SK;

public class Translator

{

    Kernel kernel;

    public Translator(Kernel kernel)

    {

        this.kernel = kernel;

    }

    public IDictionary<int, string> Translate(IDictionary<int, string> textList, string language)

    {

        var pluginDirectory = Path.Combine(System.IO.Directory.GetCurrentDirectory(), "Plugins/TranslatePlugin");

        var plugin = kernel.CreatePluginFromPromptDirectory(pluginDirectory, "Translator_" + language + "");        

        var json = JsonConvert.SerializeObject(textList);      

        if (!string.IsNullOrEmpty(json))

        {

            var output = kernel.InvokeAsync(plugin["Translator_" + language + ""], new() { ["input"] = json }).Result.ToString();

            if (!string.IsNullOrWhiteSpace(output))

            {

                Console.WriteLine(output);

                return JsonConvert.DeserializeObject<Dictionary<int, string>>(output);

            }

        }

        return new Dictionary<int, string>();

    }

}

这个类中构造函数中接收传入的Kernel对象，这个Kernel对象是指

Microsoft.SemanticKernel.Kernel  

//
// Summary:
//     Provides state for use throughout a Semantic Kernel workload.
//
// Remarks:
//     An instance of Microsoft.SemanticKernel.Kernel is passed through to every function
//     invocation and service call throughout the system, providing to each the ability
//     to access shared state and services.
public sealed class Kernel

暂且理解为调用各类大模型的Kernel核心类，基于这个Kernel实例对象完成大模型的调用和交互

另外，上述代码中有个Prompt模板文件读取的操作。

        var pluginDirectory = Path.Combine(System.IO.Directory.GetCurrentDirectory(), "Plugins/TranslatePlugin");

        var plugin = kernel.CreatePluginFromPromptDirectory(pluginDirectory, "Translator_" + language + "");    

从Plugins/TranslatePlugin目录下读取指定的KernelPlugin，例如Translator_en英语翻译插件和Translator_pt 葡萄牙翻译插件

 var output = kernel.InvokeAsync(plugin["Translator_" + language + ""], new() { ["input"] = json }).Result.ToString();

调用KernelFunction方式实现GPT4大模型调用

 //
    // Summary:
    //     Invokes the Microsoft.SemanticKernel.KernelFunction.
    //
    // Parameters:
    //   function:
    //     The Microsoft.SemanticKernel.KernelFunction to invoke.
    //
    //   arguments:
    //     The arguments to pass to the function's invocation, including any Microsoft.SemanticKernel.PromptExecutionSettings.
    //
    //
    //   cancellationToken:
    //     The System.Threading.CancellationToken to monitor for cancellation requests.
    //     The default is System.Threading.CancellationToken.None.
    //
    // Returns:
    //     The result of the function's execution.
    //
    // Exceptions:
    //   T:System.ArgumentNullException:
    //     function is null.
    //
    //   T:Microsoft.SemanticKernel.KernelFunctionCanceledException:
    //     The Microsoft.SemanticKernel.KernelFunction's invocation was canceled.
    //
    // Remarks:
    //     This behaves identically to invoking the specified function with this Microsoft.SemanticKernel.Kernel
    //     as its Microsoft.SemanticKernel.Kernel argument.
    public Task<FunctionResult> InvokeAsync(KernelFunction function, KernelArguments? arguments = null, CancellationToken cancellationToken = default(CancellationToken))
    {
        Verify.NotNull(function, "function");
        return function.InvokeAsync(this, arguments, cancellationToken);
    }

　继续封装GPT4TranslateService，构造Microsoft.SemanticKernel.Kernel 类实例。

using System.Globalization;
using Microsoft.SemanticKernel;

namespace LLM_SK;
public class GPT4TranslateService
{
    public IDictionary<int,string> Translate(IDictionary<int, string> texts, CultureInfo cultureInfo)
    {
        var kernel = BuildKernel();
        var translator = new Translator(kernel);
        return translator.Translate(texts, cultureInfo.TwoLetterISOLanguageName );
    }

    //私有方法，构造IKernel
    private Kernel BuildKernel()
    {
        var builder = Kernel.CreateBuilder();
        builder.AddAzureOpenAIChatCompletion(
                 "xxxxgpt4",                  // Azure OpenAI Deployment Name
                 "https://****.openai.azure.com/", // Azure OpenAI Endpoint
                 "***************");      // Azure OpenAI Key

        return  builder.Build();
   }
}

四、测试调用

这里我们设计了2种语言，英语和葡萄牙的文本翻译

var culture = new CultureInfo("en-US");
var translator = new GPT4TranslateService();
translator.Translate(new Dictionary<int, string>(){{ 1,"电站"}, {2,"终端不可用"},{3,"充电桩不可用"} ,
{4,"场站"},{5,"充电站暂未运营" }},culture);

culture = new CultureInfo("pt-BR");
translator.Translate(new Dictionary<int, string>(){{ 1,"电站"}, {2,"终端不可用"},{3,"充电桩不可用"} ,
{4,"场站"},{5,"充电站暂未运营" }},culture);

输出的结果

{"1":"Charging station","2":"Charging point unavailable","3":"Charging station unavailable","4":"Charging station","5":"Charging station not in operation yet"}
{"1":"Estação de carregamento","2":"Ponto de carregamento não está disponível","3":"Ponto de carregamento não está disponível","4":"Estação de carregamento","5":"A estação de carregamento ainda não está em operação"}

五、总结

以上是基于SemanticKernel和GPT4实现一个智能翻译服务的Demo和框架，大家可以基于这个示例继续完善，增加更多动态的数据和API调用，例如将JSON数据写入数据库

同时还可以记录翻译不稳定的异常，手工处理或者继续完善Prompt。

周国庆

2024/2/17