The Rise of the Chatbot

A chatbot is a computer application that performs the task of an intelligent intermediary connecting people, digital systems and the internet of things. Chatbots replace the need for a graphical user interface with a conversational experience.

Chatbots undertake a wide range of tasks. They assist people to carry out manual tasks, they provide therapy, they play the role of avatar teachers to help people develop skills such as interview techniques, and they are playing an increasing role in healthcare diagnostics.

Here we will look at the history of the chatbot, current state of the art, and what we might expect of chat boxes it in the future.

The beginnings of the chatbot

It all started with Alan Turing’s paper "Computing Machinery and Intelligence" published in 1950 which posed the question “Can machines think?”. Naturally, this continues to stimulate debate. In the paper, Turing proposed what is now known as the “Turing Test” to determine whether a computer can think like a human being. Although this has a plethora of alternative interpretations, we usually consider it to be a measure of how closely a computer can simulate human intelligence.

Many consider the ELIZA program, developed by Joseph Weizenbaum, as the archetypal chatbot. This program identifies cue phrases and words in the input and responds with pre-determined phrases. While there is no intelligence contained within ELIZA, it is still able to fool many people and, in some, even evoke an emotional response.

A version of ELIZA was programmed to simulate a psychotherapist, though PARRY, a similar program developed in 1972 by Kenneth Colby, a psychiatrist, simulated a paranoid schizophrenic. It was so successful at doing so that it fooled 52% of psychiatrists who tried it. Perhaps that isn’t surprising. These kind of chatbots are in fact quite clever in the way they repeat variations of the human input to create a sense of human presence and companionship.

Natural Language Processing – the game changer

Natural language processing (NLP), also called computational linguistics, elevated chatbot technology to the next level. Fuelled by a massive increase in affordable computing power, the availability of big linguistic data, advances in machine learning, plus linguistic research into the structure and application of human language, NLP gave birth to a new generation of chatbots. 

NLP goes far beyond the statistical approach to chatbot human interactions. Rather than relying on lookup tables and pre-programmed response choices, the technology sets out to understand, learn and create the human language. Its applications are far-reaching. Apart from powering chatbots, it also has applications in machine translation and in assisting conversation between people.

The first breakthroughs occurred during the 1980's when cracks began to appear in Noam Chomsky’s generally accepted theories on transformational grammar (TG). TG adopts a logical approach to understanding both the surface structure and underlying meaning of sentences, taking into account both syntax and context. While Chomsky’s approach provides a mathematical understanding of language, it isn’t compatible with machine learning, which leverages corpus linguistics (CL).

CL analyses language by examining large bodies of machine-readable text and derives empirical rules based on that analysis. Clearly, CL and TG are in many ways polar opposites, but that isn’t to say they don’t each have a role to play; as we mention later. However, while TG is incompatible with machine learning, corpus linguistics is entirely compatible with the way in which machines learn. The modern approach is to develop statistical models, such as the cache language model, which determines the probability of word sequences and can even deal with input errors.

Machine learning uses both supervised (annotated data) and unsupervised (non-annotated data) methodologies, with the advantage of unsupervised learning having access to all the text on the world-wide-web. Such approaches are making significant progress. Current state-of-the-art systems combine both machine learning and linguistic structure tools which together can identify semantic, contextual and syntactic information.

As an aside, as he elucidated in his book “The Blank Slate,” Stephen Pinker believes that the way in which children learn language is to develop empirical rules of syntax based on what they hear; not too dissimilar from the way in which machine learning carries out the task.

NLP and Chatbots

These advances in NLP have truly empowered chatbots to do the tasks we wish them to do. They can understand the different ways in which people speak; for instance, ask Alexa what the weather is like today in any common dialect or accent, and you will almost certainly receive the right answer. But there is much more going on under the skin. For instance:

• NLP provides to ability to translate and parse your requests into system tasks.
• Contextual abilities allow them the understand and remember contextual information.
• Emotional processing allows them to consider your emotions when responding to you.
• Action abilities enable them to carry out your request.
• Learning abilities empower them to learn continually as they interact with people.

Ever more enterprises are adopting chatbots particularly for customer service functions, marketing, and training: 40% of large enterprises state they have or will implement chatbots by 2019. This is great for business, and it can improve customer experience, for instance, there is no longer the need to wait 20 minutes for a customer service agent to answer your phone call. But, they can also be frustrating; if they are unable to deal with your query you will still end up at the end of the queue waiting to talk to a human being,

The future of chatbots

As clever as chatbots might appear today, there is still room for improvement, in particular in their Intelligence. AI is progressing apace. Deep Mind’s Alpha Zero can massively outperform previous AI iterations such as Alpha Go, which easily beat the world champion at the world’s most difficult game. Alpha Zero taught itself to play in a matter of hours having been taught only the basic rules. Applying such machine intelligence to chatbots is likely to increase their abilities substantially.

We are also likely to see higher levels of integration where chatbots and GUIs work together to optimise customer experience. For instance, the chatbot could display specific fields requiring textual input or automatically fill forms based on spoken words.

New applications will also emerge. Recently the North-eastern University in Boston, Massachusetts began a trial with people diagnosed terminally ill. They were supplied with an end-of-life chatbot to help them make their final decisions.

Chatbots are also being taught negotiating skills and how to drive a hard bargain. In tests, the chatbot was quick to learn bluffing tactics and deceit and was it almost as good at negotiating as human beings. Perhaps in the future, they will become better negotiators than humans.