Does a Bot Need Natural Language Processing?
NLP bot developers agree: “It depends.”
One of the most exciting things about the rise of chatbots is their use of artificial intelligence — especially machine learning — to mass-accomplish tasks that neither an army of interns nor an army of experts could match, and to derive wisdom beyond that of the crowd by analyzing the crowd’s billions of conversations with cold math. Yet anyone who chats with a few bots knows the frequent frustration: This thing doesn’t understand what I’m saying.
There are basically two kinds of chatbots in early 2017, while natural language processing is still learning to understand human conversational speech: Bots that risk trying to parse anything you type at them, and bots that limit your input to a few safe option buttons. Octane AI, which publishes Chatbots Magazine, currently opts for the button approach. But of course we wonder ourselves: Is that doomed?
Bot developers will tell you that it depends what your bot is trying to accomplish. In theory, a bot with a human adult level of linguistic skills would be awesome. In practice, natural language processing — NLP to anyone in the field — isn’t there yet. But in the right contexts for the right applications, NLP can make for an easier-to-use interface to features and services. Moreover, an NLP-equipped bot can give the human on the other end the feeling that they’re having a conversation, rather than poking through tedious software menus in yet another part of their lives — first it was the coffee maker, now it’s Facebook?
Marcellus Gaag is CEO of chatbot developer Sodima Solutions, which developed a chatbot with NLP for the website of a college campus political campaign (look in the lower right corner.) “The bot appeared more personal because of its NLP capabilities,” he says. Rather than try to have the bot answer every question imaginable, the campaign chose to focus on the two areas students would most want it to talk about — the campaign, and student small talk — and to integrate the bot into their website in a complementary fashion.
“Our customer was able to make their website much less content-heavy and focus it more around the call-to-actions,” Gaag says. “Students participating in the campaign said that it was ‘fun’ talking to the bot.”
How Does NLP Work?
To put it very simply, NLP software isn’t looking for keywords in your text, like a search engine. It uses knowledge of sentence structure, idioms, and machine-learned pattern recognition to try to match what you say to an “intent” which has been “classified,” which means the bot has been programmed to identify certain things people want from it, and act upon them. This involves four different areas of AI — see our What is NLP for an intro. (We’ll be explaining much more about AI and NLP going forward.)
The scope of all human intents is a lot for a bot to deal with. But your bot doesn’t need to. Gaag’s team kept it down to fewer than three dozen intents:
Whenever an intent is “classified” and used in a conversation, the bot can provide an action or quick response. This bot is trained with 20 small-talk intents (e.g. ask it: “tell me a joke”, “what are you up to?” or stupid things like “catch me outside”) and 12 specific intents that are specifically trained for this customer. These are not a lot of classified intents, but it is enough to solve a niche problem for a campaign and it gives users instant gratification.
By integrating the bot on their website, the candidates could use a web page when that was best — such as long platform statements — and let the bot serve as a sort of front-desk personality. It didn’t try to understand and answer anything and everything students typed at it, so that it could focus on the campaign and have ready answers for idle banter.
Where Does NLP Work?
Adding natural language for simple domains is overkill,
says Dennis Thomas, CTO at NeuraFlash, which develops AI tools that integrate into Salesforce. He’s a computer scientist who worked for several years with Nuance, makers of the popular Dragon series of voice-interface applications. Thomas says NLP can be a buzzkill in the wrong context: “When you have a visual medium and buttons can accomplish the task in a couple clicks (think easy re-order), open-ended natural language is not making the user’s life easier.”
Humans seem hardwired to stall at menus with more than five to seven options on them, though (a limit stressed in my old Apple user interface guide for programmers, but ignored by nearly every software maker since.) Rather than present a banking customer with 31 options at once, a chatbot can take away customers’ confusion and obscurity by asking them to tell it what they want, like a human banker behind a window in a branch office. It works in personal banking, because there are a finite number of transactions which the bank is expected to perform.
Thomas says the even bigger question is: What are you trying to do with your bot? “NLP within a chatbot for traditional self-service tasks (like your banking example) is everyone’s favorite example,” he says. “Another place where NLP is a big win is when the bot’s objective is focused on helping users with the discovery phase of products or shopping. Finding the right item via conversation (‘Do you have men’s spring jacket, either fleece or windbreakers’) helps to drive the user’s goal, as well as the product criteria to match to the company’s inventory. This discovery pattern applies to products, media (songs, movies), vacations, etc. The bot can help surface options quickly to start, driving to a call to purchase when the user’s ready to pull the trigger.”
Vivek Ramesh, an engineer and entrepreneur with several startups, says not everything works best as a bot: “A flight booking bot would ask far too many questions and this defeats the purpose of simplicity. Train booking, bus booking, flight booking, hotel booking, food ordering are best left to be dealt with by apps. NLP is best suited for cases where conversations are the way to go. For example, you ordered a product on Amazon and you have some problem with the delivered item. It is natural for you to try and talk with someone, saying, ‘Hey, I got a defective item, is this something you can fix?’ Traditionally, people would use email or live chat for this if it was available. Bots would be a solid use case for this.”
NLP also works where the user’s responses are within a narrow context or form, but potentially unlimited. The Bus Uncle Messenger bot in Singapore flat out tells you: “I will understand if you tell me bus number and bus stop code, e.g. ‘bus 64 at 60121'”. It’s a perfect job for a text-driven bot. A button interface for bus routes and times, or a tree of Web pages, is the bane of riders in many cities. Instead of needing to go to a Web page with a form from their phones, why not just let riders type route numbers into Messenger? Plus Bus Uncle’s answer won’t get lost if your phone browser’s window disappears.
Why Chat? Because People Trust It
Vasili Shinkorenko, founder and CEO of BotCube (who curates the awesome-bots list on GitHub) says NLP has proven popular with customers for expense management for an intriguing reason:
“We use a ton of personal financial management tools: apps, Google Sheets, Excel, Notes, etc. But none of them could easily remind us to track our expenses right, and on time. It could be solved using AI and bots. It’s like chatting with your friend, when he asks you how much you spent on this thing. You don’t hesitate to answer these messages, and they are the most efficient reminders and notifications, because the messaging app is a very intimate and trustful space.”
Bots you can talk with also work for dispensing advice, he says, in a way most people wouldn’t accept from a website, or a face-to-face friend. A chatbot can point out that you’re going over your monthly budget without being told to mind its own business. “People need [this advice] to make reasonable decisions,” Shinkorenko says. “What’s interesting is that people don’t think this is advertising, or that we’re trying to sell them new investment plans or banking products. It’s because of chat. Chat = friends right now, and for the next few years this will be true.”
