My very first encounter with a chat-bot happened a couple of years
back when I was trying to lodge a complaint with Amazon customer service. I
felt that I was handed out a poor delivery experience by the logistics company
that was shipping on-behalf of Amazon and got miffed. I was just trying to find
out where to register complaints in the Amazon site and got a pop-up from the
right lower corner of the web page offering a conversation with an executive. I
wasted no time in accepting the offer. After the initial pleasantries, I took
time to write a best possible statement that captured my issue along with
emotions. By any instant messaging standard, that was a big paragraph rather
than a usual bits-and-pieces chat message. The moment I hit the 'Send', I was
greeted with an even bigger response in good English explaining why and how it
could have happened, promising to fix the problem. Only 2 things were running
in my mind that time-
- The executive was extremely
fast and smart
- My problem was not that
unique and Amazon faces these issues day-in and day-out and they have
standard response templates
Even then I was slightly perplexed if someone can read my big
statement, grasp the issue, select the best possible response template,
customize as needed and send it to me in such a fraction of a second? I
continued the conversation. All my questions and chat messages were responded
in a rapid fire mode, good English and NO chatty-informal replies even though I
felt the executive's exchanges were slightly off-tracked at times. Still, by
the time, I ended up the conversation my mind was swamped with the thought - am
I talking to a human executive?
We got serious about building a bot ourselves a year ago when one of
our flag-ship customers challenged us if we can develop a virtual
conversational agent that can take the load off his customer services team and
whether they can be relieved from regular, mundane queries and service
requests. Since we were working with the customer for very long time, we were
able to choose the use case that we thought as the right one for chat-bots and
took a plunge. Even by the second half of 2016, we were able to locate some 20
bot frameworks - 4 or 5 were really good in vision even though the vision was
not matched by good implementation. We explored
Microsoft bot framework, Facebook's
Wit.ai, Google's
Dialogflow
(formerly API.ai) and some name-less frameworks that eventually faded and not
traced even by Google.
I showed special interest in Dialogflow because I thought it had
lower entry barriers (I am not responsible if you understand 'no entry
barriers' as 'credit card number not required' :)) I was stunned by its
beautiful , elegant UI/UX that abstracted complex natural language processing
algorithms (NLP). For a curious developer who didn't have any background in
advanced mathematics, statistics or computational algorithms, Dialogflow
offered a great launching pad for NLP/ML based conversational apps
Creating a Conversational Model in
Dialogflow
How do we interpret others' speeches? We need to infer-
- What's being spoken about? a
topic, a movie, a personality, a product, a selfie or any 'thing'?
- And what they are saying
about the 'thing'? good, bad, great, hopeless, want, don't want and so on
Dialogflow calls the first one as 'entities' and the second one as
'intents' (most NLP technologies call them the same). Modeling a conversational
solution involves identifying the entities and intents of the problem domain. A
complex flow may include hundreds of entities and intents. If you are modeling
a 'pizza order' bot, then different pizza varieties, sizes, toppings, thin or
deep crust, deserts are all essentially entities. As the modeler, you need to
define all the different ways in which an entity will be called (in the
Dialogflow terminology, it's called 'synonyms')
Intents could be enquiring
the menu, wanting to know the special offers and discounts, ordering the
dinner, asking for bill etc. Intents may need to be aware of the context in
which it can happen. For example, 'ask for bill' intent can happen only after
'order pizza' intent. Dialogflow lets you define the input contexts (when this
intent can possibly occur?) and output contexts (announcing to other members of
the model that something happened already). Continuing on the same example,
'order pizza' intent goes as input context for 'prepare the bill' intent
Dialogflow understands the intent by asking the modeler to give a
few samples of 'User says' (i.e. example user utterances). This can be
considered as the initial training data. For example, someone wants to know the
menu may express the same in different ways ('what you have for lunch today?',
'menu card please' or 'do you have veggie garden?'). The modeler is expected to
train the model with these sample utterances. Accuracy of these sample data
will determine the accuracy of the bot itself as Dialogflow uses machine
learning (ML). In the 'User says'
sample, you can optionally define the entities - 'lunch' could be an entity,
'menu card' could be an entity, 'veggie garden' could be an entity. You can
make some entities mandatory in some intents. Example, the user must convey
what ('entity') she is ordering in the 'order pizza' intent. When you make the
entity mandatory, the bot will expect the user to specify what pizza they want
while ordering
Confidence Score
Bot is trained with the 'User says' utterances and entities so that
it can identify the user's intent. During the initial stages, your training
data may not be adequate and bot will have difficulty in identifying the
intent. Hence it declares a 'Confidence Score' whenever it matches user's
utterance with a specific intent. As the modeler, you can configure the
threshold for confidence score. If the confidence score is below the threshold,
then bot will not match the user's utterance with the intent
Intent Response
Dialog flow lets the modeler to configure the intent responses. When
the bot is able to match a conversation with a specific intent, the bot
delivers whatever intent response configured to the user.
Integration with Databases and 3rd
Party Services
You might be curious if you don't need to be a developer for
modeling a domain. You are mostly correct and it will be suffice if you know
the domain the you are modeling. But most real-time scenarios will require some
additional business logic push in the form of data external to the bot or 3rd
party web services like a news feed from a media house or weather information
from the meteorological department. When you are faced with such a scenario,
you might need to write code to connect to the external databases and services.
Bot can connect with any RESTful service (created using technologies like
NodeJS/Express, ASP.NET MVC web API, Spring MVC etc.) For example, if the user
asks for menu, bot may need to get the available menu items from the restaurant
chain's product catalog that might have been globally defined. Github has
sample code for these web hook services
Channels
The model that you defined is mostly channel agnostic and hence the
same model can be used across multiple channels with minimal or no change to
the model. Dialogflow natively supports all the popular channels like web,
skype, slack, FB messenger, Kik, Alexa etc. You need to register your bot with
the respective channel providers. For example, if you want to create a bot that
works in Skype, you need to register and integrate the bot with Microsoft bot
framework. Once done, your bot will sit along with your Skype contacts and you
can start conversation with the bot as if you initiate a conversation with one
of your contacts
Training the Bot
Since Dialogflow uses ML algorithms, the efficiency and accuracy of
the bot increases as the days go by and you continue to amass real-time user
conversations. Dialogflow provides a rich interface for customized training of
the bot where you can identify additional entities, intents you might have not
thought of before, but the users rightfully expect your bot to respond to. Bots
may not be accurate all the time, but improves with valid training data and the
modeler may need to adjust the model based on the training data
Small-Talk
Dialogflow supports a concept called 'Small-Talk' that takes care of
normal utterances like 'how are you', 'hi', 'you are welcome' etc. You don't
need to do anything other than just enabling the feature in the console
For hosting the chat-bot for your users, you will need to create it
as a Google
Firebase project and you
may need to pay depending on different factors. There is a vibrant community
that operates both in
Github and
Dialogflow. Both Google and the community build a lot of pre-built agents. You
can refer them to get the best practices and some unique ideas. You can
export/import the bots and invite other users to develop or review the bot
using the Dialogflow console and Google accounts. Google has an extensive
technical
documentation here. Having talked
about Dialogflow, I will encourage the readers to take a look at Facebook's bot
toolkit and Microsoft bot framework/
LUIS
(Language Understanding Intelligent Service)
