Conversational Bots

The dream of humans being able to hold a natural conversation with a computer is older than the computer itself. The notion pervades science fiction, a genre where it has always been possible to converse fluently with intelligent machines. The computers of the spaceship Enterprise and C3PO are two significant instances.

Conversation agents record early successes

Current solutions are known as chatbots and already play a role in many of our lives. Companies are seeing initial success with the use of Conversational Agents, which are employed in areas such as customer service, online marketing, and banking as well as staff recruitment and onboarding processes.

The ability to automate conversations represents a great - if not a huge - commercial opportunity. Gartner predicts that “by 2021, 15% of all customer service interactions will be handled entirely by an AI, an increase of 400% over 2017." This assessment is based on observing the current market for AI - in particular, the dialogue technologies driving the transformation and automation of customer service

Chatbots Don’t Chat All That Well

While 15% of all customer service interactions seems at first glance to be a sizable chunk, the question of why "only" 15% could also be asked. If we look at the scenarios of customer service interactions, then we see a pattern of trivial tasks with complex problem solutions.

It is clear that this 15% is only dealing with a limited range of requirements. And this perfectly describes the status of chatbots today: they are not very good at chatting. Generally, chatbots lead the customer through a predefined decision tree, which really amounts to nothing more than a click path.

What Makes a Good Conversation?

Chatbots effectively act as a kind of website within a messenger service like WhatsApp or Facebook Messenger. Is this what Mark Zuckerberg meant when he said:

"We believe that users should be able to address a company the same way they address a friend."?

Surely not. A conversation is a performative act, i.e., each participant contributes the content through his or her words. Automating dialogues between humans and machines is about understanding what the customer wants to achieve in a particular situation and responding appropriately - in a personalized and contextualized way.

Natural Language Understanding Drives Development Forward

With the latest advances in the development of artificial intelligence, we have come much closer to the goal of practicable human-computer interaction. Conversational interfaces have become a big trend: Amazon Alexa, Google Assistant, and Siri are common examples.

The driver for these successful products is Natural Language Understanding (NLU). While at the beginning NLU was driven by programmed rules and grammars, technology today learns from so-called training data. No previous grammatical knowledge is required.

The Challenges faced by developers

This works really well for many suppliers of NLU products, but it is only a first step towards natural conversations with computers, as current solutions still face major challenges:

Comprehension

Most modern systems focus on understanding the user's intention in a single statement and conduct the conversation by looking at this statement in isolation. The dialogue doesn't work that way! The dialogue is referential and also takes into account previous information and details that are sometimes implicit.

Dialogue

Isolation and interpretation of individual statements without regard to the overall context lead to rigid and predictable dialogues that react to each intent in a predefined way. These systems are not able to approach anything like the "flow" of a natural dialogue – but this is what people are used to when they speak to one another.

Training

Modern approaches based on machine learning require thousands of examples before they deliver proper results. For the majority of applications, however, it is not possible to provide this quantity of material.

Remember: In most cases, dialogue behavior is predefined. There is a defined decision tree or a model that has learned from example dialogues. Unfortunately, these solutions are not dynamic: the bot does not decide what to answer based on the context which is what should happen.

So, What Does the Bot Need to Understand in Order to Deliver a Better-Quality Conversation?

Context

The chatbot should be able to act on what has already happened in the conversation as well as what is currently being discussed.

Single pairs of requests and answers become long threads, and the bot should be able to understand which options are appealing for the customer. Also, as the chatbot accumulates information, it needs to be able to use it to disambiguate user input - i.e., clearly assign conversational elements.

An example: If the user of a cook chatbot is presented with two recipe suggestions and asks: "Is the first one vegetarian?", the bot can recognize which of the suggestions "the first one" refers to. Following this, if the user next asks "And the other?", the bot should conclude that the implicit question is whether the other proposed recipe is also vegetarian.

Personalization

If the chatbot can use the knowledge it has learned about the customer, it can then offer options that are more relevant. Dialogue behavior becomes data-driven. This means that the underlying data source and the results of individual user queries are brought together. Depending on how many results a request has, the bot can adjust its response accordingly: If there are too many results, it can try to narrow the search by asking more questions. If there are too few results, it can actively extend the search.

Better Conversations Thanks to Flexible Dialogue Behavior and Appropriate Strategies

At present, architectures to meet these challenges already exist. I am convinced that at the core of future chatbots there will be a dialogue behavior that can adapt to the complicated steps of a conversation. These steps arise from the course of the conversation and its associated information - because different user stories require different strategies to achieve their respective goals. For example, an explorative search for vehicles to meet the customer’s needs can be much more open and flexible than determining whether the customer is entitled to a certain level of credit.

A one-size-fits-all approach is doomed to failure. If a human is to conduct a dynamic dialogue with a computer, it will not be the user or the bot that drives the conversation, but both together.

About the Author

Jens Kretschmann studied visual communication at the art school in Hamburg Wandsbek and gained agency experience in Hamburg, Düsseldorf, and Bielefeld. He is a specialist in user experience and accompanied projects for several DAX 30 companies such as Deutsche Post, Allianz, Mercedes-Benz, Thyssen Krupp, and Siemens. With his start-up, Mercury.ai, Jens and his co-founders have developed the interactive chatbot KIM for Nestlé’s Maggi cooking studio, which then became a benchmark for Facebook's messenger services. According to the news magazine "Wirtschaftswoche", he is one of the “100 people who influence the Internet economy.”

For questions please feel free to contact Jens Kretschmann directly.