📈

Understanding Conversational
Systems
Conversational Prelude
All of us enjoy conversations
What if learning was also conversational?
Gear up to hop on to the world of Bots and Conversational Agents . Hope you have a
good conversation ahead!

What is a Chatbot?

Understanding Conversational Systems 1

 A chatbot is defined as a program that interacts with humans through an interface
and has been designed in a way to mimic response like humans.

It is also known as a Conversational Agent or Dialogue System.

Voice-Based Assistants

I have used Google Now, Siri and Cortana. Are they also chatbots?
Yes, they are also chatbots, specifically voice-based personal assistants .

Evolution of a Bot
Science and Engineering should be taught with a little bit of history.
Ok, so how did Chatbots evolve over the years into something that we are using
today?

Web Interfaces paved the way for Mobile Interfaces. As more and more
users started using mobile apps, service providers started creating native
mobile apps.

Understanding Conversational Systems 2

 Mobile apps became saturated over a period of time as users got tired of
installing and uninstalling them. Only a few prevailed.

Apps that survived were mostly messaging apps.

Messaging apps had captured the mindshare of users. This meant that users
are available and responsive via messaging.

This lead to opening-up of multiple products and services through the chat
platform. A slew of companies started launching their respective chat
platforms.

With the advent of progress made in Natural Language Processing, it became

easier to build conversational interfaces.

Personal Bot

So now that we got a peek into the history of bots, what are the various types of
bots available?

Several types of bots are getting developed today. First, you must know
about Personal Bots .

Understanding Conversational Systems 3

 Personal Bots are messaging bots serving as a personal assistant. An
example would be a shopping bot hosted via Facebook Messenger.

These bots have a very specific focus and communicate directly with the
users on a one-on-one basis.

Domain Specific Bot

Good to know. What are the other types of Bots?

Hold your horses; you will learn them - one bot at a time. The next type of bots
is Domain Specific.

They represent single service and are very focused on the domain and the
service.
For example - Airline Travel Bot is the be-all and end-all stop companion for all
travel related queries from Booking flights, Travel Enquires to Membership
Benefits.

Consumer Bot
Ok. Are there any bots that are chatty and help us stay in touch with our
favorite items?
Yes. Some bots fit that bill. They are called Consumer Bots .

Consumer bots are

less task-oriented as Business Bots and

focus more on personal and end user experience.

For example: Mitsuku is a bot built for a Conversational purpose on the Kik
Messenger Platform.

Business Bot
Gotcha. So are all bots just consumer-facing?

You just asked the right question at the right time. Not exactly. There are several
bots for Business Process Functions.

Understanding Conversational Systems 4

 Say you have questions related to HR Policies or Regulation standards . These bots
will guide you through the nuances of the policies in the way you can comprehend
them clearly.

Voice Bots
I am tired of listening about just Chat Bots. Are there any other interesting bots?
Yes. Voice Bots. All your voice-based personal assistants come under this
category.

Google Assistant, Amazon Alexa, Microsoft Cortana, and Apple's Siri are a few
examples. They take voice commands to execute a specific task.

Super Bots

So, is there a bot of all bots?

Yup. They are called Super Bots .

They are not specific to a Domain or Business Area.

You can ask any question to them on any topic. They respond well if they can find
the answer, if not they admit that it is beyond their understanding.

For example: Google Assistant

Understanding Conversational Systems 5

 Anatomy of a Bot

Parts of a Chatbot
That was a well-animated image of a Chatbot. Could you throw more light on
the anatomy of a Bot?

Yes sure. Bot consists of the following parts.

Human

User Interface

Chat Application

Natural Language Parser

Component to Trigger Actions

Responder

Understanding Conversational Systems 6

 Human in the Loop

Why should you consider a human as a part of the chatbot?

Good Question. It is the human who triggers the conversations , and the response is
also directed towards a human . So that is why the user is a vital component of a
chatbot.

User Interface
Why do you need a user interface here?

Another important component is the User Interface (UI) .

UI acts as a medium of communication between the human and the rest of the
components.

Enough thoughts should go in building a UI that is conducive for generating a

rich experience during conversations.

Chat Application

Understanding Conversational Systems 7

 Yes. I understood the importance of the UI. Can you enlighten me on the other
part of the chatbot?
Next comes the Chat Application on which the bot is hosted. Multiple vendors
provide a platform to host bot.

Facebook, Slack, KiK, etc. are some popular platforms.

Natural Language Parser

A while back, you were mentioning about the advent of NLP so where does all
that happen?
Hold on buddy I am just about to explain that part now. There is a component
named parser that is used to extract the intent and the entity of the input.

Concepts from NLP are leveraged in this component.

Intent and Entity

Now, I get the NLP part but what is the significance of Intent and Entity?

Understanding Conversational Systems 8

 The intent and entity are used to get the context of the conversation and to
trigger the correct action relevant to the question.

Action
I got reasonable clarity on the intent and entity part. So what's next?
After extracting the intent and entity, its time for some action. Based on the
context the right action should be triggered. In some cases, it is an API call, in
other scenarios, it is a DB query.

Responder
You mentioned about the responder, what does that do?
A Responder will respond to the human, based on the response of the action.

This closes the loop.

Bot Platforms

Understanding Conversational Systems 9

 Can you give me some information on the various Platforms available for
building and hosting bots?
Yes. Before that, you should understand the various AI as a service providers and
channels.

Service providers provide a platform for you to build and sometimes host your
bots.

Once you host a bot, you create a channel for interaction between humans
and bots.

Slack - Business Bot Platform

Understanding Conversational Systems 10

 Can you please share some details about Bot Platforms?
Absolutely. There are multiple platforms available today to build and host bots.
Here, you will be learning only a few of the popular ones. Let's get started with
Slack.

Slack is a Business Bot Platform available in mobile and desktop.

You can do a variety of actions with Slack API.

Post Messages

Expose Slash Commands

Expose Clickable Controls

Event API Subscription

Administrative Actions

Consumer Bot Platform

Understanding Conversational Systems 11

 Nice. Can we now move on to the next platform?

Sure. Let's move to Facebook Messenger Platform . Facebook is a leading Consumer

Bot Platform with more than a billion active users. Facebook API provides some
really nice features.

Content Posting

Check Delivered Callback

Receiving Messages

Predefined Button Actions

Canned Responses

Sharing Geolocation Information

Voice Bot Platform

Good. Can you explain about the Voice Bot Platforms?

Understanding Conversational Systems 12

 Alexa is a popular bot from Amazon. Alexa is enabled through a device
named Echo . Customization and third-party integration are supported through the
Alexa Skills Kit.
Some of the features include:

Registering the secondary voice commands for specific actions

Receiving inputs from clients

Converse with voice back to user

Connected Home, integration with IoT devices

Legacy Platforms
Is that all or should I know about any other new platform?

Well there are many new platforms available but don't forget the legacy platforms
for communication like

Email

SMS

They are old but still used ubiquitously.

Choosing the Right Platform

Now I am spoilt for choices. How do I choose the right platform?
Good Question. Selecting the right platform is a critical step in building a bot. It is
a tricky process because it is not just a Technology, Business, or Design driven
decision. It is a combination of all of them. Some points to consider before
narrowing down on the platform are:

Audience : Who are the end users of the bot?

Use Case : B2B or B2C Scenario

Features Required : Voice Enabled or just Conversation

Device Preference

Understanding Conversational Systems 13

 This is not an exhaustive list of points to consider but will help you in choosing the
right platform.

Evaluating the Platform

How could one evaluate the platform?
Build a Proof of Concept (PoC) and get users to converse through the platform. This
is a good way to evaluate your platform.
Just to draw a parallel correlation,

Without habitat, there is no wildlife. It’s that simple. — Wildlife

Habitat Canada

Use Case: Conversational Commerce

Understanding Conversational Systems 14

 That was enlightening information on platforms. I now get the hang of things
about bots. I want to understand more about the Use Cases. Can you help with
that?
Let's start with a popular use case, Conversational Commerce .

Consumers can use a bot to add items to their shopping basket.

Bots can be used for booking tickets and guide you on your travel itinerary.

You can also book a cab with the help of your Uber or Lyft bot.

Use Case: Business Functions

Please go on. Enlighten me more on these use cases.

Now You will learn about some Business Functions .

After the advent of the DevOps way of working, collaboration and

communication are inclining towards ChatOps, i.e., Managing DevOps via
Chat.

Legal Bots are reviewing Non-Disclosure Agreements and sharing inputs with
the Legal Team.

Bots are getting leveraged for Corporate Workflow Management.

Productivity and Coaching Applications

Are Business Bots similar to Productivity improvement and Coaching?
Not exactly. For Productivity Improvement actions like:

Reminders

Creating to-do lists

specific bots are getting developed.

Coaching is a big area where bots are being leveraged.

For example: Personal bots are being used as trainers for weight loss.

Understanding Conversational Systems 15

 Alerts and Nudges
A while back, you spoke about E-Mail and SMS. Those are age-old means of
communication. Can bots perform similar functions?
Absolutely. Bots cannot completely replace some traditional modes of
communication instead could aid in nudging and alters functions like

you can get a notification from a bot after you have successfully completed a
financial transaction.

You can get reminders from bots on some specific goals that you have set and
not progressed.

They act as nudging agents in these scenarios.

Intermediaries
Alright. You have given me some bots related use cases. I am tired of listening
to similar examples. Are there any exciting use cases where bots are getting
leveraged?
Yes. As the human in the loop, bots are now acting as intermediaries between
humans.

A ride-sharing bot will connect one human to another.

Travel bots help connect a group of people who are working towards a similar
goal like Mountain Climbing or Deep Sea Diving.

FAQ
That's interesting. I have always felt that the Frequently Asked Questions
(FAQs) could be enhanced with the help of bots. Am I right?
Yes, you are getting there my friend. FAQs are getting replaced by bots for
a better customer experience. Here, bots are acting as the first line of support for
all customer queries.

Artificial Intelligence

Understanding Conversational Systems 16

 So, I have had a perception that Bots are a manifestation of AI. Am I right?
Not exactly. AI is just a significant component of a Bot and not the essence of a Bot . You
are not alone my friend, many have had the same misconception. I am here to
declutter all that.
So what exactly is Artificial Intelligence(AI)?

According to Wikipedia, "Artificial intelligence (AI) is intelligence exhibited by

machines. In computer science, an ideal 'intelligent' machine is a flexible rational
agent that perceives its environment and takes actions that maximize its chance
of success at some goal".

Bots and AI
Gotcha. Can you give more insights on this AI part?

AI is not the one-stop shop for building bots; instead, AI is a set of tools and
techniques that can be leveraged to build a conversational system.

Natural Language
Now that I got some clarity on the AI part, let's get into the Natural Language
aspects.
Sure. All conversations happen through Natural Language

NLU

Understanding Conversational Systems 17

 What is Natural Language Understanding (NLU)?

NLU is a series of steps that take the input text and extract
the intent and entities of the input.

Training

Understanding Conversational Systems 18

 Natural Language is unstructured, how do I make my bot efficient at extracting
intent and entities?
Machine Learning is used extensively to train bots at identifying the various
aspects of Natural Language. Most of the NLP tools come with pre-trained
models.

Some Examples
Can you give me some examples of input text?
Say you are conversing with a Travel Bot. You could say,

"I want to buy a flight ticket to New York", or

"I want to get flight tickets to New York", or

"Let us get a flight ticket to New York".

All the three sentences have the same intent, just that they are worded differently.

Grammar

Understanding Conversational Systems 19

 So, do I have to train my bot on Grammar as well?
To a certain extent yes. NLP has a technique called Part of Speech Tagging which
helps understand the construct of the input better.
This, in turn, helps in extracting the entities and intents to perform the desired
action.

Conversation Management
Ok. I have a fair idea on AI and NLU now. What next?

After understanding the Intent and Entities, it is now time to take action and
respond to the input.

Dialogue agents have two-way communication, so you have to map a correct

response to the input.

Can you throw some more light on the Conversation Response?

Managing the conversation has a good amount of AI involved.

AI component understands the context and renders the correct response

based on the input.

When there is a lot of context switching, humans handle it better, but

machines find it cumbersome.

Context Switching
Ok. So what are the hardships in switching context?
One word might have different meanings in different contexts.
For example,

Java is a programming language and a coffee variety.

Python is a snake and a programming language.

Getting the context is very important to render the correct

response.

Understanding Conversational Systems 20

 Semantics
So this is all about getting the semantics right, is that correct?

Absolutely. Conversation Management as a field is still at a very nascent stage. A lot of

progress is required to be made in this field.

Decision Dilemma
From our previous conversations, I get the confidence that I can develop a Bot
even without AI. Am I right?
Absolutely.
AI is not the de facto component in all bots. You should use it as the situation
demands.

Using AI
In what scenarios should I use AI?
Scenarios that demand

High use of Natural Language Understanding

Frequent Context Switching

Predictive Analytics

Sentiment Analysis

Image Recognition

require the use of AI.

Using AI
Got it. So from what I understand, situations where there are many
unstructured components, AI can be used. Am I right?
Yes. Where there is a lot of complexity in the input, AI can help reduce the
complexity.

Understanding Conversational Systems 21

 Not to use AI
What are the scenarios where I need not use AI?
Scenarios where bots

Send Financial Reports

Perform News Aggregation

Send reminders and alerts

do not require any use of AI

Takeaway
So what is the takeaway from this AI conversation?
Well, you must say that if you ask me then

AI is very expensive so use it cautiously

AI can sometimes complicate your conversations

so exercise caution before taking a call on AI.

Conversation Types
That was a lot of information on Bot. Now I want to understand more about
conversations. What are the types of conversations that one can have with a
bot?
Ah! So now you want to get into the weeds. Conversations can be broadly
classified into

Task Led

Topic Led

Task Led
Can you throw some light on Task Led Conversations?

Understanding Conversational Systems 22

 As the name suggests, these conversations mostly happen with an objective of
accomplishing a specific task. Completing a stipulated task will be the expected
outcome of the conversation.

Example
An example would give more clarity. Can you share one?
Say you are using a bot to order coffee. The successful outcome of conversation
would be taking the order of a coffee.

Coffee Example
Cool. So if I just converse with a bot saying give me coffee, is the task
accomplished?
It is not as easy as you think my friend. People across the globe have a variety of
preferences to order a cup of coffee. So each preference has to be modeled as a
flow-through its entirety.

Modeling Flow
Gotcha. So what exactly happens during each flow?
I think you need some coffee first to be more attentive. Ok, let me not wait until
then.
During each conversation

the Entities are extracted

based on the entity, the intent is mapped to each of them.

Sometimes what happens is the conversations go out of context. It is very

important to bring them back to the right context.

Topic Led Conversations

Can we move to topic led conversations?

Understanding Conversational Systems 23

 Absolutely. Unlike task led, topic led conversations are long. It depends on the
users' engagement with a specific topic.
So does a chatbot become a chatterbot in topic led conversations?

In a way yes. Topic led conversations can go in multiple directions. They do not have
a well-defined end state. They are more designed to keep the customer engaged
on specific topics.

Example
Again, I am used to learning by examples. Can you share one?
Sure. Say you are conversing with a bot on a movie you recently watched or on
some company you have invested in the stock market. You can converse on
different dimensions on the same topic.

Takeaway
So what I understand is that Topic led conversations are mostly used for user
engagement, and task led conversations are used for business scenarios. Am I
right?
Awesome. High Five. You are getting there.

Interacting with Bots

So, is text the only mode of interaction with Bots?
Not exactly, the text is one of the modes of interaction. Some of the ways of
interacting are

Files

Audio

Video

Emojis

Buttons

Understanding Conversational Systems 24

 The list goes on...

Files

How can files be used as a means of interaction with Bots?

Say you are interacting with a Tax Filing Bot. Instead of entering all the fields in
your tax filing document, you can just upload the file. Since the data is structured,
it can be easily parsed for further processing.

Files
Now it makes perfect sense. When should I leverage this feature while
designing a bot?
Good Question. Files are used in scenarios where a lot of information has to be
processed, and your chat window is small to accept all that information in one go.
You can also use them when you have tables.

Understanding Conversational Systems 25

 Audio
Can you throw some light on Audio Interactions?
Sure. For personal assistants like Google Assistant, Siri, Amazon Alexa, etc. the
mode of interaction is Voice. This type of interface is called Vocal User Interface .
You trigger commands like Ok Google "How is the weather in New York?" and you
get a response back with the required details.

Audio
That was useful. But how do we address the accent issue with voice interface?
Good question. Voice might not be a useful interface in every scenario. Voice is
recommended only for private interactions. A lot of training is required to get the
bot to understand various accents.

Video
Now I understand how Audio interactions happen. I am really intrigued about
videos. Can you help?

Sure. Video interactions happen in the following scenario. Say you are interacting
with a bot to troubleshoot a specific problem and you need an approach. Your bot
can respond with a video that will help you troubleshoot the problem. This is one
of many scenarios where videos can be used as a means of interaction.

Emojis
Somewhere above, you did mention emojis as means of interaction? Is that the
way forward in terms of communication?
Not exactly. Emojis are more of reactions to conversations. These reactions can
also be used to take actions, acknowledge, and respond.

Emojis

Understanding Conversational Systems 26

 Can you give me some example of how Emojis are used for conversations?
Sure. Look at the image above. The same conversation can be reacted with emojis
and expressed with the help of emojis.

More on Emojis

Understanding Conversational Systems 27

 Is that all? Any interesting examples?
Yes. Look at the image above. You can create polls and get a response from
emojis. You can also aggregate the response and perform analytics using those
numbers.

Buttons

Understanding Conversational Systems 28

 That was insightful. Can you now talk about Buttons?
Yes. Buttons are a wonderful way to navigate a conversation. You can confine the
user to a pre-defined set of responses and guide the user in the desired direction.

Buttons

Understanding Conversational Systems 29

 So should I use buttons only for canned responses?

Not just for Canned Responses. It can also be used for other scenarios. Look at
the image above. You want to collect orders based on some predefined inputs.
You can leverage buttons in those scenarios.

Context

Understanding Conversational Systems 30

 I have been intrigued by this thought on how a bot understands the context.
Can you throw some light on that?
Roger that. Say you are in a conversation with a Travel bot. You are chatting about
the Eiffel Tower. The conversation goes as shown in the image above.
In the above conversation, it refers to Eiffel-Tower the context of the discussion.
This can be achieved with sufficient training on varied topics to understand the
context.

Context

Understanding Conversational Systems 31

 Got it. Although there was never a mention of the Eiffel Tower in the following
questions, the word it specifies the context. Am I right?
You are getting there my friend. Let us take another example. Look at the image
on the top. You will get more clarity on the intent and context.

Context and Intent

I saw the travel booking example in the previous card, but I am still unclear on
how to train a bot on the context and intent. Can you help to clarify?
Sure. Some of the variables in the conversation have limited scope like start and
end date; some have a global scope like User Information. These variables are
tightly coupled with the intent and context of the discussion.

Context and Intent

Why are Context and Intent so Important?

Bots are normally used to a request to response setup. They should not time out
after multiple conversations and should remember the context from the beginning.

Understanding Conversational Systems 32

 Context and Intent

I still feel there is an overlap between Context and Intent. Can you please help
me clarify?
Look at the Image above. Hope you got the answer to your question.

Getting the Context

Are there some tips and tricks to get the context right?
Yes. You can get the context from

Pronouns

Understanding Conversational Systems 33

 Words like "it", "this", "hers", and "his" are pointers for getting the context.

Rich Controls

Memory
I got the hang of what context and intent are for a Bot. How different is a
memory?
Memory is generic when compared to context. Here, a bot should be able to pull
intents at various time steps and take actions.

Amnesia
Do bots suffer from Amnesia?
If you do not train them well, then they can suffer from Amnesia.

Closing Thoughts
That was very informative. Now I have some clarity on Context and Memory.
Very Good. Getting a bot to understand the context and memory is one of the
most challenging parts in designing a bot. It is also a good area of growth from a
technical standpoint.

Human In the Loop

So far, I have understood multiple aspects of building bots and enabling
conversational experience. Is there anything additional I have to know?
Yes. You should understand a very important aspect of building conversational
systems i.e. Human-in-the-loop.

Humans or Bots

Understanding Conversational Systems 34

 Ok. So, are humans going to work with bots or bots going to replace humans?
To say the truth, it is a little bit of both.

HITL
I am getting confused now. Can you help me understand this concept?
Sure. The steps to handle Natural Language Understanding are Understand -> Act
-> Respond.

Let us take an example: You ask a bot

“What’s the status of SE 123?”

Decomposing the above question

Understand: User wants to understand the flight status of Flight SE 123.

Act: Query Database or Hit API to find the flight status.

Respond: Reply based on the output received from the previous step.

Understanding Conversational Systems 35

 Having a human intervention in any of the above steps is called Human-In-The-
Loop

Human Intervention
When should Humans Intervene in this process?
Humans should intervene when

Algorithms find it difficult to understand new scenarios

Algorithms have no clue how to perform a certain action

Unsuccessful actions

Input is out of scope

Best Practices
Are there any best practices for this scenarios?
Well, nothing is written in stone, but a few good practices can make things better if
implemented correctly.

Data Collection should be done properly

Gauging possible outlier scenarios and include them during training

Handling exception scenarios should be thought out well.

Understanding Conversational Systems 36