Creating Digital Assistants
The technology allowing organizations to create digital assistants is here and available today.
The function of a digital or intelligent assistant is being delivered today. Siri, Cortana, Alexa, Samsung Bixby, and Google Assistant are examples. You have heard about them, and many have likely experienced them. They are usually called personal assistants. But what about a company-specific digital assistant at a bank, insurance company, airline, or retail store?
The company-specific digital assistant was the topic of a recent presentation, "Creating an Effective Company Digital Assistant" delivered by Dr. William Meisel at the SpeechTEK conference that took place at the end of April. Meisel is a consultant, editor, and publisher of LUI News, which covers commercial developments using human speech or text. This blog summarizes his SpeechTEK presentation.
What is a Digital Assistant?
The key to today's natural language processing is to limit the context of user input that the automated processing has to understand. For example, if you are a bank offering credit cards, you should expect questions such as, "What's my account balance?" and "When is my next payment due?" Other examples of user phrases would be, "I need a loan," or "I'd like to finance my mortgage."
All types of organizations should also expect questions such as, "What can I say?" and may have to decide what to do about off-topic questions like, "Are you a person?"
You can help the user understand what kind of things they can say. You could tell a customer calling a credit card call center, "Please state what you are calling about. You can ask about your balance or a payment." The digital assistant interacts with the caller to provide the information or action requested. The type of interaction can depend on the platform you're using. If you build a bot for Facebook, the interaction will be by text. If you build a skill for Amazon's Alexa, the interaction will be by speech. The core natural language understanding technology developed for text is largely applicable to speech and vice versa.
Not Just for Consumers
A fundamental question is whether this is another fad that will die away. Is the technology ready for prime time? Gartner says that by 2020 customers will manage 85% of their relationship with an enterprise without interacting with a human. Meisel estimates that sales of over $600 billion globally would be generated for companies by interactions through digital assistants of all forms.
There are channels by which the user interacts with a digital assistant. These can be through a website with a text chatbot or a microphone icon. One can reach the same assistant on a mobile phone through a browser or the connection can be through a dedicated app that the user launches. Or, on a PC, the digital assistant can be a direct application as opposed to working through a website. If the customer is calling into customer service, the natural language interaction can be through an IVR system, which automates the call without a frustrating series of layered options.
Some of these channels emphasize text interaction and some voice interaction. As previously noted, the core natural language processing is very similar with both speech and text inputs.
Speech input requires speech-to-text processing and can introduce recognition errors in the text that the natural language processing must interpret.
Text input can be cheaper to develop and use, but has its own problems. The processing must handle misspellings, typos, and the shorthand sometimes used in text messaging. An advantage of text chat is that users associate it with text messaging, and don't expect an instant response, as they might with a speech interaction, so the speed of response is less of a problem, part of why it is cheaper.
The core task is to understand what intents a user might have in interacting with the digital assistant. "Intents" is actually a technical term used by most of the natural language tools. Specifying intents requires an understanding of your business more than some theoretical analysis. Given an utterance or text entry by the user, you reduce that general context with a specific intent, a narrower context. For example, is the customer calling about an account balance, questioning an item on their statement, or indicating a desire to transfer money between accounts? Each of these is an intent that will be handled differently.
How About Tools
Facebook provides a tool called Wit.AI. Once set up, it identifies an intent and retrieves the variables required to complete that intent.
Amazon's Alexa has a full platform in the cloud and free deployment for moderate use. The company provides tools for building the skills, and recently announced improvements.
Amazon also gives you the option of using the core technologies behind Alexa. Amazon Lex Services let you build a branded specialized digital assistant without working through Alexa, giving you direct contact with your customer at a usage-based price for both voice and text.
IBM provides cloud services you can use through APIs. They have what they call the Watson Virtual Agent that can be used to create a chat bot. According to IBM, this does not require anyone to have machine learning experience to use.
Another example is Artificial Solutions' Teneo Platform, which is a full platform with all the desired capabilities. Finally, Interactions has a platform that also supports natural language assistants.