The History of Chatbots

built to be skimmable for the busy reader

built to be skimmable for the busy reader

Chatbots have evolved from pre-scripted replies to self-thinking and creative rebuts over 52 years. Revenue for the Speech and Voice Recognition Software Developers industry is expected to increase an annualized 17.0% to $12.6 billion over the five years to 2017 (IBIS World, Report OD4531).

End-Goal

Have a human that they are talking to another human when they are interacting with a machine program. The Turing test, developed by Alan Turing in 1950, is a test of a machine’s ability to exhibit intelligent behavior indistinguishable from that of a human (Wikipedia). You may have heard of Alan Turing from The Imitation Game where in 1939 British MI6 Intelligence began cracking the Nazi Enigma cipher machine.

Origins of Chatbots (1960s)

The concept of a chatbot began in 1966, when Joseph Weizenbaum created an early natural language processing computer program, ELIZA. The program was designed to mimic human conversations by matching user prompts to pre-scripted responses.

ELIZA: the first chatbot was not the best conversationalist.

Hard Coding Chatbot Responses (1980s)

Up to the 1980s, most systems were based on complex sets of hand-written rules. Often with psychologists leading early projects, researcher would build a tree of possible conversations and type out pre-written responses. The chatbot would then reference the available text strings to pick a response to provide the user. Over the course of the next 30 years developments added personality, conversational consistency, and natural language processing abilities.

Hard Coded Responses: a decision tree that an early chatbot would reference to determine what to say next.

What is Natural language processing?

Natural language processing (NLP) is the study of computer systems that can interpret speech and text — either spoken or written. Within a chunk of text, NLP will look for three aspects to which it attempts to understand:

  • Semantic: Identifying the meaning of a specific word is difficult when words can have multiple meanings. A phrase like “Let’s meet in the Winter Garden” can be confusing to the system unless it understands that the Winter Garden is a term used by business students at the University of Michigan. If the system assumes that a garden is where plants grow, it won’t provide valuable insight.

  • Syntax: Sentence or phrase structure varies widely in every day use. Take the example: “Professor Dittmar joined the students already with NLP experience.” Who has the experience: the Professor or students? If the system assumes incorrectly, it won’t provide valuable insight.

  • Context: Context is the concept being discussed. If a student describes a global experience in London as “sick” does that mean that the students visited the hospital frequently or it was better than expected? If the system assumes incorrectly, it won’t provide valuable insight.

In Action: Natural language Processing

Putting it all together, let’s look at the sentence “Professor Wu shot the paper across the University.” Taken separately, the three types of information would return results along the lines of:

Semantic information: person — act of moving quickly — collection of ideas — place of learning
Syntax information: subject — action — direct object — indirect object
Context information: this is about a Professor sharing ideas with peers.

Together these ideas paint a picture of what is being described. To make it ever more difficult, we humans keep changing the meaning of words. NLP has to follow our slang, abbreviations, and nicknames.

“The language doesn’t take a vacation, and neither does the dictionary. The words we use are constantly changing in big ways and small, and we’re here to record those changes.” — Emily Brewster, associate editor at Merriam-Webster

How does NLP evolve to stay up-to-date with every day use?

The answer is artificial intelligence (AI). Most AI work now involves machine learning (ML) because intelligent behavior requires considerable knowledge, and learning is the easiest way to get that knowledge.

AI = teaches systems to do intelligent things
ML = teaches systems to do intelligent things that can learn from experience
NLP = teaches systems to be intelligent, learn from experience and understand human language

Artificial intelligence is an umbrella term that encompasses machine learning and natural language processing.

Present Day Chatbots (2000s)

Can we use chatbots daily? Yes. Chatbots are being integrated into anything that has a microphone and speaker. With the computing conducted in the cloud, the on-device hardware can be basic.

Verto Analytics is an industry leading smart assistant research firm.

Siri

By far, the most oft-cited reason for using a personal assistant app (a use case claimed by 71% of respondents) was to ask a question or search for something. Users aren’t accepting of Siri’s ability to perform this action well. Between May 2016 and May 2017, Siri lost 7.3 million monthly users (nearly 15% of its total).

Take-away: Siri had the early advantage and is losing users and marketshare.

Alexa

Meanwhile, Amazon Alexa usage has been skyrocketing — jumping 325 percent in monthly active users — that is, from 0.8 million to 2.6 million monthly users, as its user engagement also increased from 10 percent to 22 percent during the same time frame of May 2016 to May 2017.

Amazon is dominating the Smart Speaker market. Using Prime Day as a mechanism to distribute product through promotion, they have a stronger channel into the home than Google.

Amazon’s Skills are apps for the smart assistant. The number of Skills available in the United States grew from 6,000 to 24,000 Skills December 2016 to December 2017. Similar to Apple in 2008 to 2010 when they were supporting the app store developers, Amazon has announced in March 2018 that they will enable a library of audio sounds to developers in order to spur activity on the platform. This is an arms race to see which platform (Amazon vs. Google) can achieve higher customer and developer adoption.

Take-away: Alexa is winning on three fronts (1) smart speaker market share, (2) developer contributions, and (3) user frequency of engagement.

Google Home

The device market share above points out a key trend: price point. The Echo Dot (March 2016) was released years after the original Echo (November 2014), however it is quickly gaining market share and is anticipated to become the most popular Echo device.

Google experienced the same trend as the Home (November 2016) was released before the Mini (October 2017) and is now the more popular of the two devices.

Take-away: The Google Home Mini ($49) is gaining adoption faster than the higher priced Google Home ($129). Google has doubled-down on their efforts to encroach on Amazon’s early lead.

Slack

Users can enter regular sentences to a chat-bot on Slack. Slack has hundreds of developer created chatbots. If a user typed “show me conferences in March”, the Slack service recognizes the intent to query for event data. Next, it understands that the search is by date. Furthermore, it translates “March” into a date range from 2018–03–01 to 2018–03–31. Utilizing Conversation dialog actions, the search request is routed as a query. The result data is post-processed and a text answer is returned to the Slack user.

The Slack App Directory includes a section “Bots” with 100s of apps for free installation.

Bots for Messenger

In the wake of the Cambridge Analytica scandal Facebook quietly announced Monday it is pausing its app review process, which means developers are no longer able to launch new apps or chatbots on the Facebook ecosystem. That abrupt halt, even if temporary, is a thudding blow to any app developer who had hoped to debut a new experience on Facebook this week.

Imagine hundreds of hours of work, tens to hundreds of thousands of dollars in investment capital, and dozens of clients disappearing at any given moment at the whim of a few lines of code” - Troy Osinoff, cofounder of digital agency JUICE, via Facebook post on March 28th, 2018.

Facebook Messenger appears to be a reasonably promising channel for distribution. With over 1.2 billion monthly active users, Messenger has enough of an audience to matter to marketers. And it has more than 100,000 bots and just as many developers as of April 19, 2017.