Chatbots are exploding. Their growth mirrors the widespread adoption of APIs and apps and new, innovative bots are launched almost every day. In fact, odds are you’ve interacted with one already, whether you know it or not.
That said, even with the deserved hype around them, a lot of folks don’t know how to train and improve their chatbots. In fact, some chatbots are little more than search engines, sending customers to help articles and FAQs that don’t really answer their specific questions.
At CrowdFlower, we’ve seen a lot of bot projects. We understand how to train a chatbot so it can hold real conversations to help real people solve real problems. We’ve put together some FAQs and all our relevant resources on this page. Here are a few to get you started:
- The Data Science Behind Making Sure Your Chatbot is Worth Talking To (free eBook)
- Getting a Chatbot to Understand 100,000 Intentions
- How Yalo used CrowdFlower to Create a Spanish Language Chatbot
- Free Chatbot Training Dataset
What is a chatbot?
Simple! A chatbot is a program that can carry on a conversation with a person. Chatbots, sometimes called conversation interfaces or virtual assistants, are usually powered by rules or AI and generally lives in any major messaging application (think Slack or text messages).
What are chatbots used for?
The most common use case for a chatbot is in customer service, answering common questions for real-world users. Other chatbot uses include product recommendations, personal shoppers, giving life advice, or even bots for just talking to or playing games with. Since the most common (and probably the most valuable) use case is customer service, we’re going to spend most of this page talking about them.
Why build a chatbot?
For B2C companies, there are a two major benefits to building a chatbot. Namely:
- Cost savings: Simple online chat functionality really helped streamline customer service practices. Agents who use chat to solve problems tend to solve them much more quickly than they would over the phone and, additionally, they can juggle several conversations at once. Chatbots speed this up even further. Once your chatbot is properly trained, it can answer myriad requests simultaneously, as well as farm out the complex issues it can’t quite understand to people on your support staff.
- Visibility: The easier it is to find–and interact with–your company, the better. Since chatbots are often built on messaging sites like WhatsApp or Facebook, that means you’ve got a virtual agent ready to solve problems where your customers are already. That’s convenient for them and shows that your company is truly invested in ironing out their issues. It also increases engagement with your customers on common social channels, which is always a good idea.
How do you program a chatbot?
Generally, chatbots are trained with four major kinds of data projects. We can handle them all at CrowdFlower. And while we explain each in depth in our free chatbot ebook, here’s a quick overview:
- Utterance collection: This is a data collection job where you ask people to help come up with various ways to ask for the same thing (or make the same generally utterance). This trains your bot to understand that “gotta change flight” is the same as “can you help me reschedule my trip tomorrow?”
- Relevance: Here, you want to train your bot to give relevant responses by making sure the replies your model thinks are correct are in fact relevant to real user questions.
- Intent: Similarly, here you want to judge exactly what the user’s intent is when they’re engaging with your bot. This is closely tied to relevance, generally.
- Entity recognition: A more complicated natural language processing (NLP) training mechanism that helps your bot understand nuance and context. For example, “do you have any 9s” is different if your bot’s talking about shoes than if it’s talking about golf clubs.
Why do chatbots fail?
Chatbots often fail because they’re overly ambitious. In other words, they try to solve too many problems.
It’s much smarter to identify the most common issues your customers face and solve those with your bot first. This helps your living, breathing customer service agents handle more nuanced problems and not feel like they’re doing the same thing day in, day out. Not only that, but zeroing in on a few key genres of support means that you can train your chatbot to be super effective clearing up those issues instead of trying to be all things to all customers. If you deal with a lot of shipping issues, for example, start by honing your bot to be an expert in that exact problem. Then, expand to other, less essential (but still important) areas.
What’s the right tone of voice for a chatbot?
Chatbots should mirror your brand persona. If your company is playful, your chatbot should be too. But nobody wants their banking chatbot tossing out “lol”s after a balance inquiry either.
What’s more important is that you understand how a chatbot should converse. Chatbots should be curious. Because conversation is hard. We all say things differently and seemingly identical issues can lead to immensely different chats. Your chatbot should ask just as many–if not more–questions than it answers. It should make sure it fully understands the issue before it attempts to solve it. And asking questions can really help get it there.
Chatbot vs. human: what are the advantages of each?
First off, don’t think of it as human vs. bot. It’s really about making sure both excel, which they can, in tandem. Maximize each approach by parcelling out the most appropriate issues to the most appropriate agent. People are great at nuanced conversations with a lot of back and forth, especially conversations that evolve or have a tendency to move in unexpected directions. Chatbots, on the other hand are great at speed and simpler conversations that can more easily mapped out.
The added bonus? Your people spend more time using their creativity and expertise while avoiding rote issues easily handled more quickly (and without frustration) by a bot.
Want to get started?
If you’re interested in creating or training your own existing chatbot, please don’t hesitate to reach out. We’d be happy to help show you all we know.