Photo by Akshay Chauhan on Unsplash

Chatbots: From Unstructured Data to Conversation

Have a Conversation with your Customer

Cobus Greyling
7 min readAug 14, 2019


Any conversation via a conversational interface, like a chatbot, poses the challenge of creating structure from a highly unstructured medium. Which is in this case, a natural language conversation.

Break it down to the basics

Chatbot which is Contextually Aware

I always find a good approach to Natural Language Understanding application design, is to break the process down to its most basic form.

Any conversation is constituted by the following elements:


The basic notion that any conversation has a specific context within which that conversation takes place.

Even when humans have a conversation, context needs to be established first. If not, it leads to misunderstandings.

Context within a conversation is the parts of the conversation that immediately precedes the current passage or words, allowing for clarity and meaning.

Context assists both the user and the conversational Interface to construct meaning from the conversation.


Conversations, and especially story telling was the first form of virtual reality; reproducing events of the past and bygone realities. Customers interacting with the conversational Interface might have contacted us in the past.

It is important to identify those customers and present continuity to their conversation. This is in line with any human conversation; when we meet with a familiar face, we expect a new conversation. but also, entertain the possibility that we will continue a previous conversation.


Again, in any conversation data is captured by one or multiple parties. One or more parties have certain objectives (data) they would like to capture with the conversation.

This leads to an informal process whereby form filling takes place; but in a conversational manner.

Directed Dialog

Any conversation is constituted by dialog, an interchange of words, phrases and ideas. In most cases the dialog is directed by one or more parties. The direction of dialog is crucial for any conversation so one or more parties can extract meaning, intent and information.

IBM Watson Assistant Contextual Entities Test Environment

Directing dialog allows for conversations to have a natural narrative arch. Where an introductory phase is followed by rising action and data collection.

Once the bot has reached its unpronounced objectives, there is a falling in action until resolution from the bot’s perspective. But more importantly, the resolution on the user’s side is more a sense of closure and comfort.


Conversations between humans take place via various modalities or mediums. Conversations can be in person, video calls, voice calls, text etc.

The same holds for a chatbot; the bot can have the same anthropomorphized attributes which is available via different modalities and interfaces. some of these modalities include:

Web chat, Facebook, Twitter, WhatsApp, WeChat, Telegram, Smart Speakers (Digital Assistants), Google Home, Amazon Echo, Show, Spot, Dot (Alexa), Apple HomePod.


I believe the medium impacts the message. So, even though the current functionality and persona is replicated on various mediums, the presentation will differ. This relates to the media options in terms of buttons, marques, images etc.

Hence in some cases the user interface will have to rely more on Natural Language Understanding. Apart from text, other supported media types are documents, images and audio.

Contextual Entity Extraction

Service Design Fulfillment

Any product manager should be meticulous when it comes to Service Design, as it is ultimately the differentiation that will attract a client to your brand, as opposed to competitors and keep them on board.

Service design determines the outcome of Customer Experience (CX). CX is the way a customer or prospect feels after interacting with your interface.

This sensation the customer has after interacting with your brand is to a degree very subliminal, but indeed very possible to craft, hone and reproduce.

From here a blueprint can be created and the following elements can be addressed:

Support Systems

What is the companies “North Star”? In brief, the North Star of a company focuses on the product and the core value it delivers to customers. The value customers receive from the product is measured by key steps in the consumer life cycle. Very often, the North Star is related to the acquisition of new customers and the retention of existing customers.

Form Filling via Conversation

Net promoter Score (NPS) an assist in defining the achievement within the North Star vision.

The value that the product creates for customers drives engagement and ultimately creates value and strategic direction for your company.

Putting down a concept to test should not be discarded for meeting a deadline.

The concept should be tested, and an iterative approach should be followed.

Other key elements are:

· Back-end systems
· Key Metrics
· Delivery Targets
· Customer Centric
· Technical feasibility
· Holistic approach

Rasa Tensorflow Pipeline; Language Independent

Success Criteria — Definition

As a team, you want to design and create a service which is co-created by key employees and clients, and that connects to people on an emotional level, thereby developing into a service they love, that is interdependent and interrelated.

Functional Fulfillment

The initial functional fulfillment can be limited with focus on people interacting with the chatbot.

The chatbot interface can be customized for web, Facebook, Twitter, and text (SMS). Where SMS and twitter are completely text based; fully based on Natural Language Understanding (NLU) and Tone and Sentiment analysis.

Nudge Theory (Sans Dark Patterns)

Nudge theory is opposed to dark patterns. It is the methods, often subliminal, used to nudge or move a customer into a certain direction.

Negative sentiment and general complaints should be detected (tone and sentiment) moved out of the public domain (i.e. social platforms) and into a more confined and private environment, like Twitter Direct Message or Facebook Messenger.

If a call back can be scheduled it makes the experience is so much better. This nudging the client into a direct private messaging environment mitigates negative fallout by isolating the conversation.

But also holds the illusion of exclusivity and importance. Dark patterns must always be avoided, this is where the client might feel tricked, trapped or misled. Regardless of it being real, perceived, intentional or unintentional.

Multiple Contextual Entities per Intent

The Nature of Conversation

A chatbot is a conversational user interface (Conversational UI) allowing your customer to have a free-text conversation with your organization and where elements are addressed, like:

~Natural Language Understanding
~Extracting meaning and intent from the utterance
~Analyzing tone and sentiment
~Responding in a natural language format.

For any chatbot or automated assistant to work, any conversation musty have the following elements:

· Context
· Continuity / Vertically & Horizontally
· Directed Dialog

In addition to the above, the following must always be at the forefront of the solution design:

· Context: Context must be established to understand what the background, framing and setting of the conversation is. Even among human-to-human conversations a basic framing of the conversation is important.

· Continuity: We as humans can have a conversation in the morning. And then reinitialize the conversation in the afternoon, this might be referred to as vertical continuity. But the conversation could have been in person the morning, and the afternoon over the phone, or via text/SMS. Hence the continuity is horizontal across modalities.

· Directed Dialog: As humans we do not start a conversation with “ask me anything”. We have an objective, and a frame of reference for a conversation. We direct the dialog of the conversation to achieve the objection or premise on which the conversation was initiated. The chatbot has a frame of reference in which it can assist the customer. It is important that the chatbot understands where within that reference to direct the conversation.

Chatbot Persona: Anthropomorphized Chatbot

When it comes to anthropomorphizing (to attribute human form or personality to things not human) a Conversational UI such as Amazon Echo (Alexa) is a very good example. Google on the other hand, decided not to anthropomorphize its conversational interface in Google Home.

We as humans prefer the anthropomorphised approach. It is advisable that any organisation opts for this approach by having a distinct persona and name for the chatbot interface. Thought and design is required to define the persona, so it is distinct, special and a very recognizable personality.

Modalities of Conversation

A conversation can be via different modalities. A text chatbot can be deployed across modalities in terms of:

· Webchat
· Facebook Messenger
· Twitter

Later in the life-cycle, modalities like a speech interface can be introduced.

IBM Watson Assistant~ Annotate & Create Contextual Entities

Voice and Text Considerations

In brief, for later in the life-cycle, should voice be introduced, the following considerations are important:

Voice requires the adding of:
~ASR (Advanced Speech Recognition)
~Text to Speech or also referred to as Speech Synthesis.
~Voice responses must be shorted that text responses.
~Voice responses have not persistence; as in the case of text; and disappears.
~The user cannot refer back to it.
~With voice, there will be instances where a breakout is required to a more suitable modality.
~API/Integration Response Times

Photo by Mark Boss on Unsplash



Cobus Greyling

I explore and write about all things at the intersection of AI & language; LLMs/NLP/NLU, Chat/Voicebots, CCAI.