Analyzing Sentiment of Your Emails with Azure Text Analytics Service

Analyzing Sentiment of Your Emails with Azure Text Analytics Service

Azure Machine Learning allows data scientists to build and deploy predictive models. I am currently reading Predictive Analytics with Microsoft Azure Machine Learning, which in my opinion is a great resource to get started with ML. If you are a developer and don’t really want to invest in learning ML, you can use Azure ML web services published by Microsoft and other publishers in the Cortana Analytics Gallery. There are several interesting APIs, such as speech, face recognition and computer vision, available that you can use in your applications. Today, I will use Text Analytics Service, which is one of Azure ML web services available in the Cortana Analytics Gallery, to build an Outlook add-in that parses the subject of an email and classifies the email as one of 😄 😐 😠

How Text Analytics Works

In a nutshell, if you give a piece of text to the Azure ML Text Analytics service, it returns a score between 0 and 1 denoting overall sentiment in the input text. Scores close to 1 indicate positive sentiment, while scores close to 0 indicate negative sentiment. You don’t need to train the model before use, as the model is already trained. You can read more about how the algorithm works here. However, in a nutshell, this is how this service works:

  1. Obtain a large dataset of text with sentiment scores (must be easy for Microsoft).
  2. Split text into words and apply stemming (converting word to its root form) e.g. fishing, fisher or fished is reduced to fish.
  3. Create features from words. Some of the key features used are:
    • N-Grams: Generate all possible combinations of n consecutive words e.g. for “we are learning ML” and n=2, the sequence would be “we are”, “are learning”, “learning ML”.
    • Part-of-speech tagging: It is the process of identifying words belonging to a particular part of speech. A simplified form of this is identification of words as nouns, verbs, adjectives etc.
    • Word embedding: It is the process of mapping syntactically similar words close to each other e.g. car and bike are closer to each other than are car and office.
  4. Once the features have been identified, the classifier is trained with the features.

Before You Start

Not many people know that Office 13+ supports add-ins (previously called Apps for Office). You can build and debug add-ins in Visual Studio and submit them to the marketplace or distribute them privately. I encourage you to read more about the Office add-ins platform here.

Secondly, you would need to sign up for the Text Analytics service here. You would also need a key to access the API which you can download from here. Before you get started, you can play with the API on the demo console available here.

Finally, to prepare your Visual Studio environment, you would need to install Office Developer Tools which you can find here. You should also have an Office 365 account or a free developer subscription to debug your add-in.

Source Code

The entire source code of the application is available on GitHub.

In a very short time, we will have an Outlook add-in ready that analyzes the sentiment of the subject line of an email and displays an emoticon representing the sentiment. The solution has the following structure (the important files are highlighted).

HappyMailFinder Solution
HappyMailFinder Solution - Full Image
  • HappyMailFinder is the project that contains the application manifest file. You should run this project to debug your add-in.
  • HappyMailFinderWeb is a web project that contains web pages that are hosted inside Office client applications.

Go Time

In your solution, create a new App for Office project (might get renamed).

AppForOffice Project Wizard
AppForOffice Project Wizard - Full Image

In the next step, select your app type as Mail.

AppForOffice Project Wizard App Type
AppForOffice Project Wizard App Type - Full Image

On the next screen, select the options that make sure your app appears every time someone opens an email or an appointment to read.

AppForOffice Project Wizard App Appearance
AppForOffice Project Wizard App Appearance - Full Image

In the solution that unfolds, navigate to AppRead > Home.html and replace the markup inside <body> tag with the following HTML code. This markup defines the appearance of our add-in.

<div id="content-main">
    <div class="padding">
        <p><strong>Hi, I analyzed the subject line of your mail!</strong></p>
        <table id="details">
                <th>Subject Sentiment Score:</th>
                <td id="subject">Calculating...</td>
                <th>Your Mail Type:</th>
                <td id="mailType">Calculating...</td>

To invoke the Text Analytics service and display the sentiment scores in the web page that we modified, navigate to AppRead > Home.js and modify the displayItemDetails function and also add two more functions.

function displayItemDetails() {
    var item = Office.cast.item.toItemRead(Office.context.mailbox.item);
    var encodedSubject = encodeURIComponent(item.subject);
    getScoreAsync(encodedSubject, displayResult);

function displayResult(responseData) {
    var score = JSON.parse(responseData).Score;
    if (score<0.4) {
    if (score > 0.65) {

function getScoreAsync(encodedSubject, callback) {
    var xmlHttp = new XMLHttpRequest();
    var theUrl = "" + encodedSubject;
    xmlHttp.onreadystatechange = function () {
        if (xmlHttp.readyState === 4 && xmlHttp.status === 200) {
    }"GET", theUrl, true);
    xmlHttp.setRequestHeader("Authorization", "Basic YOUR BASE64 ENCODED KEY");
    xmlHttp.setRequestHeader("Accept", "application/json");

The mechanism is pretty straightforward. The function displayItemDetails retrieves the subject of the email and invokes the function getScoreAsync. The function getScoreAsync in turn sends an HTTP GET request, authorized with your base 64 encoded key, to the Text Analytics service and sends the result to the callback function displayResult. The callback function displayResult then prints the output. It is fairly reasonable to assume that scores less than 0.4 denote negative sentiments and scores greater than 0.65 denote positive sentiments.

That’s it. Compile and execute the program. You would be asked to enter your credentials to access Outlook in Office 365. Send some test mails to your account and be amazed 😄


Following is what I got from my development efforts.

Sad Mail
Sad Mail - Full Image
Happy Mail
Happy Mail - Full Image

Now it’s your turn. Why don’t you extend this application and store the sentiments in a persistent store? May I also suggest displaying a small graphic on how the sender has been behaving with you (or others)? There’s so much you can do with this service. Let your imagination loose. Hope you had fun reading this post! I will see you soon!

Did you enjoy reading this post? I can notify you the next time I post something. Subscribe.

Your 2 cents

Profile Picture

Hi! I'm Rahul Rai, an author, a programmer, and a technophile. I'm a Consultant at Readify, Sydney, Australia.

Stay up to date with interesting posts like these. Take a moment to subscribe!

Microservices with Azure now available on Amazon.
Connect with Azure Service Fabric experts and developers on LinkedIn.