A featured contribution from Leadership Perspectives: a curated forum reserved for leaders nominated by our subscribers and vetted by the Construction Tech Review Advisory Board.

Capgemini

Why it is Not a Good Idea to Build Your Chatbot in a Few Minutes

Just the other day we came around an ad on social media that went along the lines of “build your own chatbot in 10 Minutes”. Here is why this might be fund and a good idea for personal use but not valid for corporate scale deployment.

Let’s first outline the basics: a chatbot processes either spoken or written natural language as input. Based on this input a logically corresponding conversational output must be generated. This implies that conversations are designed around certain set of rules. A very basic rule, for examples, is to give a friendly response to any utterance of a friendly greeting at the beginning of a conversation. Without the need to add any further logic the response can be pre-defined (i.e. any utterance of “Hi”, “Hello”, etc.).

To enter a structured conversation more context information on two levels is needed:

• Level 1: The intent (= topic or subject) of the conversation. It needs to be clear what the user is talking about to look in the right box for answers (e.g. are we talking shoes or socks, life or car insurance?).

• Level 2: The entities (= information items) of the intent. This will be the information the bot is looking to receive from the user for a specific intent to provide meaningful answers (e.g. blue sporty sneakers size x with laces for casual events, married healthy non-smoking European male age 45 with 2 kids and a 200k EUR loan on the house)

Check This Out: Top Chatbot Solution Companies
 

Interpreting the right intent might be easy, but offering an efficient and meaningful conversation flow to distil all required entities from a user can be tricky as every new piece of information needs to be interpreted on the basis of all information already received. An efficient conversation implies a “memory” function that avoids asking questions two times. A meaningful conversation implies that the entire information content is qualified and classified before every reply. Not an easy task, if you consider bots are exposed to humans which do not follow any common interaction structure, some mean more than they say some say more than there is meaning, some jump across topics others focus on the details.

You need to train the system to go deeper in a conversation and respond contextually so that you don’t have a frowning user sitting in front of a chatbot wondering, what just came out of there?

But let’s consider the interpretation piece on the input side can be solved. How to inject the information content and orchestrate the delivery of “smart” answers as output? For this several aspects have to be considered:

• Integration into corporate knowledge / data bases: tap into your FAQs & databases

• Conversational architecture and orchestration: a very time consuming, yet important part. For this a conversational architecture is needed. It takes plenty of hours to develop the many layers of a conversations with all ways it can go. This is the base of every rule based chatbot. It involves a lot of if and else ifs and follow up intents and fall back intents. And this takes time - a lot of time. You need to train the system to go deeper in a conversation and respond contextually so that you don’t have a frowning user sitting in front of a chatbot wondering, what just came out of there.

• Calling of external APIs to tap information from the web: There are plenty of APIs out there, that can be tapped to make your chatbots a great companion. Think of a chatbot, that let’s you know whether the food on your plate is a healthy option. How you are wondering? Combine the chatbot with an image classifier trained for food and call a food fact and nutrition API to get the info you need

• Adding image classifiers to add a visual component and let the chatbot see (user identification, age verification, product recognition: Just like in the previous example: Add more cognitive functions to the bot. Make it hear and see to help you get a better experience

So as much as chatbots are fun to build within a couple of minutes-if one wants a serious companion that supports your business, a complex conversational architecture is needed to make a bot really smart.

And there a chatbot becomes much more human, because without learning, we also can only understand the question, but don’t know how to answer.

The articles from these contributors are based on their personal expertise and viewpoints, and do not necessarily reflect the opinions of their employers or affiliated organizations.

Weekly Brief