Build a Hamilton Song Recommendation SMS Bot with Machine Learning

const { TextAnalyticsClient, AzureKeyCredential } = require("@azure/ai-text-analytics");
const WhichX = require("whichx");
exports.handler = async function(context, event, callback) {
        let twiml = new Twilio.twiml.MessagingResponse();
        const key = context.AZURE_KEY_HAMILFILM;
        const endpoint = context.AZURE_ENDPOINT_HAMILFILM;
        const textAnalyticsClient = new TextAnalyticsClient(endpoint,  new AzureKeyCredential(key));
        const input = [
            event.Body
        ];
        const songs = {
        "non-stop": {
            desc: "Do not assume you are the smartest in the room. Time to do some work.",
            link: "youtube.com/watch?v=_YHVPNOHySk"
        },
        "wait for it": {
            desc: "Good things take time. You will get praise and validation soon. If you're doubting yourself, just keep going. You are inimitable, an original.",
            link: "youtube.com/watch?v=ulsLI029rH0"
        },
        "schuyler sisters": {
            desc: "Girl power! You are empowered and thus empower others. Keep your siblings close. You may be looking for a significant other, a friend, a peer, or a general mind at work.",
            link: "youtube.com/watch?v=UeqKF_NF1Qs"
        },
        "dear theodosia": {
            desc: "you get all emotional and teary over your kid or your pet. They are cute, young, innocent, and have their whole lives ahead of them, which you will make better.",
            link: "youtube.com/watch?v=TKpJjdKcjeo"
        },
        "story of tonight": {
            desc: "You are feeling emotional over what you, your friends, and your family will do in the future. The night is still young. You can all do so much and change the world!",
            link: "youtube.com/watch?v=3vqwrepaMR0"
        },
        "my shot": {
            desc: "life is tough but you are tougher. All you need is one chance, one shot per se, and here is the motivation you need to accomplish anything.",
            link: "youtube.com/watch?v=Ic7NqP_YGlg"
        },
        "alexander hamilton": {
            desc: "You don't get the hype over Hamilton. Hamilton may be new to you. This song will sum it up succinctly for you and you'll learn some history too.",
            link: "youtube.com/watch?v=VhinPd5RRJw"
        },
    };
    
    const sentimentResult = await textAnalyticsClient.analyzeSentiment(input);
    let sentiment, pos, neg, neutral, max;
    
    sentimentResult.forEach(document => {
        console.log(`ID: ${document.id}`);
        console.log(`Document Sentiment: ${document.sentiment}`);
        console.log(`Positive: ${document.confidenceScores.positive.toFixed(2)} Negative: ${document.confidenceScores.negative.toFixed(2)} Neutral: ${document.confidenceScores.neutral.toFixed(2)}`);
        document.sentences.forEach(sentence => {
            sentiment = sentence.sentiment;
            console.log(`Sentence sentiment: ${sentiment}`);
            pos = sentence.confidenceScores.positive.toFixed(2);
            neg = sentence.confidenceScores.negative.toFixed(2);
            neutral = sentence.confidenceScores.neutral.toFixed(2);
            var obj = {"positive": pos, "negative": neg, "neutral": neutral};
            max = Object.keys(obj).reduce((a, b) => obj[a] > obj[b] ? a : b);
        });
    });
   
    //Build our Bayesian model
    var whichfw = new WhichX();
     whichfw.addLabels(["non-stop", "wait for it", "schuyler sisters", "dear theodosia", "story of tonight", "my shot", "alexander hamilton"]);
    Object.keys(songs).forEach((s) => { whichfw.addData(s.toLowerCase(), songs[s].desc) } );
    const song = whichfw.classify(event.Body); 
    const reasonWhySong = songs[song].desc;
    const link = songs[song].link;
    twiml.message(`You seem to be feeling ${max}. ${reasonWhySong} We recommend listening to ${song} right now: ${link}`);
    callback(null, twiml);
};

Wow, that's a lot of code. Let's break it down.

We import Azure AI Text Analytics and WhichX at the top with:

const { TextAnalyticsClient, AzureKeyCredential } = require("@azure/ai-text-analytics");
const WhichX = require("whichx");

Then we make our Function asynchronous to give the Function more time to analyze the inbound SMS input, make a MessagingResponse object that we will later return as an outbound SMS, create variables referencing our Azure endpoint and key environment variables, and pass them to textAnalyticsClient . Lastly, we pass in the inbound text message body to an array input .

exports.handler = async function(context, event, callback) {
        let twiml = new Twilio.twiml.MessagingResponse();
        const key = context.AZURE_KEY_HAMILFILM;
        const endpoint = context.AZURE_ENDPOINT_HAMILFILM;
        const textAnalyticsClient = new TextAnalyticsClient(endpoint,  new AzureKeyCredential(key));
        const input = [
            event.Body
        ];

Next we make the key-value object holding the collection of Hamilton songs the user can be classified as. Each song has a brief corresponding description the classifier will attempt to match to according to the inbound SMS.

const songs = {
    "non-stop": {
                desc: "Do not assume you are the smartest in the room. Time to do some work.",
                link: "youtube.com/watch?v=_YHVPNOHySk"
        },
            ...
};

Now we call our client's analyzeSentiment method, which returns a SentimentBatchResult object, and create some global variables.

const sentimentResult = await textAnalyticsClient.analyzeSentiment(input);
let sentiment, pos, neg, neutral, max;

Iterate through the list of results, and print each document's ID and document-level sentiment (analyzes the whole text) with confidence scores. For each document, result contains sentence-level sentiment (analyzes just a sentence) along with confidence scores (percent confident the model is that the sentiment is positive, negative, or neutral) and more information that we do not need for this post. Lastly, we find the key (positive, negative, or neutral) that has the highest confidence level value.

sentimentResult.forEach(document => {
        console.log(`ID: ${document.id}`);
        console.log(`Document Sentiment: ${document.sentiment}`);
        console.log(`Positive: ${document.confidenceScores.positive.toFixed(2)} Negative: ${document.confidenceScores.negative.toFixed(2)} Neutral: ${document.confidenceScores.neutral.toFixed(2)}`);
        document.sentences.forEach(sentence => {
            sentiment = sentence.sentiment;
            console.log(`Sentence sentiment: ${sentiment}`);
            pos = sentence.confidenceScores.positive.toFixed(2);
            neg = sentence.confidenceScores.negative.toFixed(2);
            neutral = sentence.confidenceScores.neutral.toFixed(2);
            var obj = {"positive": pos, "negative": neg, "neutral": neutral};
            max = Object.keys(obj).reduce((a, b) => obj[a] > obj[b] ? a : b);
        });
    });

Finally, we build our naive Bayesian classifier, using it to classify the inbound text according to Hamilton songs by adding the labels of the Hamilton songs we want to classify. You could build a classifier in a variety of ways, but this is a succinct way to do so.

 //Build our Bayesian model
    var whichfw = new WhichX();
    whichfw.addLabels(["non-stop", "wait for it", "schuyler sisters", "dear theodosia", "story of tonight", "my shot", "alexander hamilton"]);
    Object.keys(songs).forEach((s) => { whichfw.addData(s.toLowerCase(), songs[s].desc) } );
    const song = whichfw.classify(event.Body); 
    const reasonWhySong = songs[song].desc;
    const link = songs[song].link;
    twiml.message(`You seem to be feeling ${max}. ${reasonWhySong} We recommend listening to ${song} right now: ${link}`);
    callback(null, twiml);