Skip to content

Latest commit

 

History

History
502 lines (337 loc) · 22 KB

README.md

File metadata and controls

502 lines (337 loc) · 22 KB
Raw

Information and user-action bot, scenario in Java

Overview

Description

This repository contains a proof of concept materials and Java code examples that were used in the creation and implementation of information and user-action bot (QnA Maker + Wit.ai) in Generali Greece.

The business challenge was to automate the work of a Generali Greece employee (underwriter) and save the time & resources during information exchange with the broker.

Challenges:

  • Implementing QnA Maker service by creating a Q&A bot answering questions of brokers towards the underwriters of Generali Greece.
  • Answers of these questions were in a CSV file with the specific area of topics around health insurance risk assessment, exemptions etc. (Questions->Answer file).
  • Natural language understanding (NLU) was required in Greek so we used another service than LUIS, Wit.ai.
  • Channel for the bot to be deployed was the custom UI of an internal Java web application that is used for communication between brokers and underwriters.
  • Not understandable question is assigned to the underwriter directly.

Solution Base Overview

The target solution and its architecture has various services implemented, resulting in a fully automated information exchange channel. At the moment, the solution includes implemented QnA Maker service and the primary method of automating tasks in the internal tool (previously supported by the underwriter and now by Wit.ai).

Below are descriptions of solution architecture (in an advanced and basic scenario):

1#92C955 The broker starts an application, and his credentials are validated.

2#92C955 Broker starts a dialog with a bot by typing a question or a command/task.

3#92C955 The question is processed by a Bot Service algorithms and passed through the specific methods of implemented services. There are a few examples of service incorporation.

4#92C955 These algorithms can invoke a method from NLU service (like Wit.ai or LUIS) to automate a task within application's backend (in this example it was a generated storage for a document upload).

4#FA8B04 Alternatively, invoke a method from Cognitive Services, like QnA Maker - to receive an answer regarding expert knowledge around health insurance risk assessments, exemptions, and so forth.

4#C80018 Or, invoke a method that search for an entity from the database, like the current status of some deal, or the date of the contract, etc.

4#E12296 In case of a specific kind of query, which is not understandable, the question is assigned to the underwriter...

5#E12296 ... who responds in person.

5#92C955 Otherwise, the bot is sending the actions/answers, done by particular services, back to the broker.

1#92C955 The broker starts a dialog with a bot by typing a query/question into the application chat window.

2#92C955 A controller is processing input and passing it through different methods and services.

3#92C955 QnA Maker, which is a service with a Knowledge Base and answers around health insurance - the specific stuff.

3#E12296 Wit.ai, which supports an NLU in the Greek language, and can automate specific kind of tasks in place of the underwriter.

4#92C955 After scoring, the controller updates an application chat window view by providing the right answer or setting up specific resource for the broker.

The flowchart below shows the idea of the scoring process:

In that case, the score is a value of confidence in each service.

Table of Contents

Prerequisites

Azure Subscription

Various ways to create an Azure Account with a subscription.

QnA Maker resource

QnA Maker - https://qnamaker.ai

  1. To create a service, go to Azure Portal and search for "QnA Maker":

  2. After you choose it, read the description and click Create:

  3. Configure your service by naming it, selecting right pricing tier (details about pricing - here), etc.:

Wit.ai Account

wit.ai - https://wit.ai

  1. Log in with GitHub or Facebook account:

Java Environment

  • Java SE 6 / 7
  • Java UI Framework
  • Documentum

Usage

QnA Maker

  1. Follow the instructions in the QnA Maker Portal in "Create a knowledge base" view:

  2. Publish Knowledge Base and get requeired Keys and Endpoint addresses (after Publish, go to Settings tab):

  3. Code snippets below represent an implemented basic methods for QnA Maker service:

QnAMaker Service Client (Remember to provide a valid Keys and Endpoints) :

public class QnaMakerKnowledgeBaseRestClient {

  private static final Logger logger = LoggerFactory.getLogger(QnaMakerKnowledgeBaseRestClient.class);

  private static String baseURI = "https://westus.api.cognitive.microsoft.com/qnamaker";

  private static String host = "https://<qnamakerservice-name>.azurewebsites.net/qnamaker";

  private static String subscriptionKey = "<from-azure-portal--go-to-qna-maker-service-resource-management-keys>";

  private static String endpoint_key = "<endpoint-key>";

  private static String primarysKnowledgeBaseKey = "<kbid>";
  
  ...

}

Knowledge Base Create method:

public JSONObject create(JSONObject createJSONObject) throws Exception {

    JSONObject restOutput = null;

    try {
      HttpClient httpclient = HttpClients.createDefault();
      URIBuilder builder = new URIBuilder(baseURI + "/v4.0/knowledgebases/create");

      URI uri = builder.build();
      HttpPost request = new HttpPost(uri);
      request.setHeader("Content-Type", "application/json");
      request.setHeader("Ocp-Apim-Subscription-Key", subscriptionKey);

      StringEntity reqEntity = new StringEntity(createJSONObject.toString());
      request.setEntity(reqEntity);

      HttpResponse response = httpclient.execute(request);
      HttpEntity entity = response.getEntity();

      if (entity != null) {
        String entityString = EntityUtils.toString(entity);
        restOutput = new JSONObject(entityString);
      }

    } catch(Exception e) {
      logger.error("Error while running QnaMakerKnowledgeBaseRestClient.create() with input: " + createJSONObject.toString(), e);
    }

    return restOutput;
}

Get Endpoint keys from a published Knowledge Base:

public JSONObject getEndpointkeys() throws Exception {

    JSONObject restOutput = null;

    try {
      HttpClient httpclient = HttpClients.createDefault();
      URIBuilder builder = new URIBuilder(baseURI + "/v4.0/endpointkeys");

      URI uri = builder.build();
      HttpGet request = new HttpGet(uri);
      request.setHeader("Ocp-Apim-Subscription-Key", subscriptionKey);

      HttpResponse response = httpclient.execute(request);
      HttpEntity entity = response.getEntity();

      if (entity != null) {
        String entityString = EntityUtils.toString(entity);
        restOutput = new JSONObject(entityString);
      }
    } catch(Exception e) {
      logger.error("Error while running QnaMakerKnowledgeBaseRestClient.getEndpointkeys() ", e);
    }

    return restOutput;
}

Get an answer from question method:

public JSONObject getAnswers(String kb, String question) throws Exception {

    JSONObject restOutput = null;

    try {
      URL url = new URL(host + "/knowledgebases/" + kb + "/generateAnswer");
      String answers = getAnswersPost(url, question);
      if (answers != null) {
        restOutput = new JSONObject(answers);
      }
    } catch(Exception e) {
      logger.error("Error while running QnaMakerKnowledgeBaseRestClient.getAnswers() with kb: " + kb + " and question:" + question, e);
    }

    return restOutput;
}

HTTP Service helper method for Get an answer from question method:

private static String getAnswersPost(URL url, String content) throws Exception {
    String restOutput = null;

    try {
      HttpsURLConnection connection = (HttpsURLConnection) url.openConnection();
      connection.setRequestMethod("POST");
      connection.setRequestProperty("Content-Type", "application/json");
      connection.setRequestProperty("Content-Length", content.length() + "");
      connection.setRequestProperty("Authorization", "EndpointKey " + endpoint_key);
      connection.setDoOutput(true);

      DataOutputStream wr = new DataOutputStream(connection.getOutputStream());
      byte[] encoded_content = content.getBytes("UTF-8");
      wr.write(encoded_content, 0, encoded_content.length);
      wr.flush();
      wr.close();

      StringBuilder response = new StringBuilder();
      InputStreamReader inputStreamReader = new InputStreamReader(connection.getInputStream(), "UTF-8");
      BufferedReader in =new BufferedReader(inputStreamReader);

      String line;
      while ((line = in.readLine()) != null) {
        response.append(line);
      } in .close();

      restOutput = response.toString();

    } catch(Exception e) {
      logger.error("Error while running QnaMakerKnowledgeBaseRestClient.getPrimarysKnowledgeBaseAnswers() with url: " + url + " and content:" + content, e);
    }

    return restOutput;
}

Get an answer from question primary method:

public JSONObject getPrimarysKnowledgeBaseAnswers(JSONObject question) throws Exception {

    JSONObject restOutput = null;

    try {
      URL url = new URL(host + "/knowledgebases/" + primarysKnowledgeBaseKey + "/generateAnswer");
      String answers = getAnswersPost(url, question.toString());
      if (answers != null) {
        restOutput = new JSONObject(answers);
      }
    } catch(Exception e) {
      logger.error("Error while running QnaMakerKnowledgeBaseRestClient.getPrimarysKnowledgeBaseAnswers() with kb: " + primarysKnowledgeBaseKey + " and question:" + question, e);
    }

    return restOutput;
}

Wit.ai

  1. To create an app in Wit.ai, click on the PLUS symbol:

  2. Add name and description and choose desired language (important!):

  3. Start by adding more and more examples of sentences, and edit/add entities (intents, Wit.ai specifics or your entities):

Testing

Here is a code fragment of an auto-reply method, which invokes a code from implemented QnA Service Client.

private Dialog autoReplyToBroker(GeneraliPrincipal principal, Contact sourceContact, Contact targetContact, String relatedEntityType, String relatedEntityCode, String dialogType, String message) {
  try {
    Dialog dialog = null;
    if (DialogTags.TASK_ASSISTANCE_ANNOUNCEMENT.getCode().equals(dialogType)) {
      QnaMakerKnowledgeBaseRestClient client = new QnaMakerKnowledgeBaseRestClient();
      String jsonString = "{\"question\":\"" + message + "\"}";
      JSONObject restInput = new JSONObject(jsonString);
      JSONObject restOutput = client.getPrimarysKnowledgeBaseAnswers(restInput);
      JSONArray jsArray = ((JSONArray) restOutput.getJSONArray("answers"));
      Double finalScore = 0.00;
      String finalAnswer = "";
      if (jsArray != null) {
        if (jsArray.length() > 0) {
          for (int i = 0; i < jsArray.length(); i++) {
            JSONObject jsObject = (JSONObject) jsArray.get(i);
            String answer = (String) jsObject.get("answer");
            Double score = (Double) jsObject.get("score");
            if (score > finalScore) {
              if (answer != null) {
                answer = Converters.asTrimmedString(answer);
                finalScore = score;
                finalAnswer = answer;
              }
            }
          }
        }
      }
      if (finalAnswer.length() > 0) {
        dialog = notifyBrokerForComment(principal, sourceContact, targetContact, relatedEntityType, relatedEntityCode, dialogType, finalAnswer);
      }
    }

    return dialog;
  } catch(Exception e) {
    throw new DataAccessException("Error while running taskService.autoReplyToBroker", e);
  }
}

The goal was to receive an answer with details about contact to Generali Greece. You can see the results of the actions on the following animations:

  1. Simple query, by providing a keyword "sms".

  2. Complex query, by providing a sentence in greek language.

Another achievement was to test a Wit.ai app by providing sentences in the Greek language:

  1. Here Wit.ai recognizes an intent by locating the name of sickness and type of action.

  2. The second example shows that Wit.ai can recognize a phone number entity and a type of action.

  3. Created entities for that test.

Learnings

QnA Maker learnings

The QnA Maker documentation shows that the service can support many languages, but in practice, it is entirely different. Here you can read more about the supported languages: https://docs.microsoft.com/en-us/azure/cognitive-services/qnamaker/overview/languages-supported

Supposedly, the language is automatically recognized from the content of the data sources being extracted. Once you create a new QnA Maker Service and a new Knowledge Base in that service, you can verify that the language has been set correctly. More details about the whole process you can find here: https://docs.microsoft.com/en-us/azure/cognitive-services/qnamaker/how-to/language-knowledge-base

🐛 🐛 🐛 In practice it is always English 🐛 🐛 🐛

I was trying to fix that in many different ways...

  • by uploading Question/Answer file before KB creation
  • by uploading Question/Answer file after KB creation
  • by providing Questions and Answers manually
  • by creating different resources with a different pricing tier
  • by creating a standalone Azure Search resource and trying to modify it and connect it with the existing resource of QnA Maker

...unfortunately without any success.

Here you can investigate my steps:

  1. Langauge support analyzer investigation in testkb index resource.

  2. Langauge support analyzer investigation in specific index resource.

  3. Creating new Azure Search resource.

  4. Updating keys and names in App Service.

Wit.ai learnings

Unfortunately, the Greek language in the Wit.ai service is still in beta, so the recognition of such things as contact is difficult to implement.

  1. Here you can see my manual selection of a name "John", but still without any improvement after clicking "Validate".

The nature of the Greek language

With these difficulties (above), we are also unable to determine how the algorithm behaves when we include Greek-language specific letters or words. The endings of words or accents that occur with individual letters are still binary values that can be understood by the machine differently.

For example brokers by typing a query can be in a hurry, and they will type

  • εχω

instead of

  • έχω

which means 'I have'.

In that particular case, the meaning is the same, with or without a tone/accent. However imagine words with a different meaning, after erasing such tone or accent.

Credits

  • Antonis Apergis - Software Development Supervisor, Generali Greece
  • Christodoulos Rokos - Senior Software Engineer, Generali Greece
  • Borys Rybak (who fell in love with Greece) - Software Engineer, Microsoft

Helpful Materials

QnA Maker:

Bot Builder SDK (Java):

Pricing:

QnA Maker Management App Service Azure Search Limitations
Experimentation Free SKU Free Tier Free Tier Publish Up to 2 KBs, 50 MB size
Dev/Test Environment Standard SKU Shared Basic Publish Up to 4 KBs, 2 GB size
Production Environment Standard SKU Basic Standard Publish Up to 49 KBs, 25 GB size