Smarter Conversations. Part 1 - Sentiment

This post starts a series of short articles on building smarter conversations with Microsoft Bot Framework. I will explore detecting sentiment (part 1), keeping the dialog open-ended (part 2), using a simple history engine to help the bot be context-aware (part 3), and recording a full transcript of a conversation to intelligently hand it off to a human operator.

Affirmation

Imagine the following dialog:

User >> I’m looking for screws used for printer assembly
Bot >> Sure, I’m happy to help you. 
Bot >> Is the base material metal or plastic?
User >> metal
Bot >> [lists a few recommendations]
Bot >> [mentions screws that can form their own threads]
User >> Great! I think that's what I need
Bot >> [recommends more information and an installation video]

It’s not hard to train an NLU service like LUIS to see a product lookup intent in the first sentence. A screw would be an extracted entity. Following a database lookup, the bot then clarifies an important attribute to narrow the search down to either plastic or metal screws and presents the results.

The highlighted sentence that follows is a positive affirmation. It is not an intent that needs to be fulfilled, not an answer to the question asked by the bot. And yet it presents an opportunity for a smarter bot to be more helpful, act as an advisor.

Sentiment

Text Analytics API is part of the Microsoft’s Cognitive Services offering. The /text/analytics/v2.0/sentiment endpoint makes it a single HTTP request to score a text fragment or a sentence on a scale from 0 (negative) to 1 (positive).

I decided to make the expressed sentiment look like an intent for my bot and so I built a custom recognizer:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
const request = require('request-promise-native');

const url = process.env.SENTIMENT_ENDPOINT;
const apiKey = process.env.SENTIMENT_API_KEY;

module.exports = {
recognize: function (context, callback) {
request({
method: 'POST',
url: `${url}`,
headers: {
'Content-Type': 'application/json',
'Ocp-Apim-Subscription-Key': `${apiKey}`
},
body: {
"documents": [
{
"language": "en",
"id": "-",
"text": context.message.text
}
]
},
json: true
}).then((result) => {
if (result && result.documents) {
const positive = result.documents[0].score >= 0.5;

callback(null, {
intent: positive ? 'Affirmation' : 'Discouragement',
score: 0.11 // <-- just above the threshold
});
} else {
callback();
}
}).catch((reason) => {
console.log('Error detecting sentiment: %s', reason);
callback();
});
}
};

Context

Now I can attach a dialog that would be triggered when the bot detects an affirmation and no other intent scores higher. The default intent threshold is 0.1 and that’s why a detected sentiment is given 0.11.

1
2
3
4
5
6
7
8
9
10
11
const sentiment = require('./app/recognizer/sentiment');

bot.recognizer(sentiment);

bot.dialog('affirmation', [
function (session, args, next) {
// ...
}
]).triggerAction({
matches: 'Affirmation'
});

Unlike other intents, however, a detected sentiment is not enough to properly react to on its own.

The bot needs to understand the context to properly react to an affirmation or discouragement expressed by a user. The bot also needs to be able to handle an interrupted dialog if an affirmation (or an expression of frustration) came in the middle of a waterfall, for example.

I will get to it in part 2. Stay tuned.