In this post, I’m addressing the second portion of unanswered questions from my webinar at the Virtual Summit on Intelligent Agents and Chatbots. You can read answers on the first portion of questions here.
You can watch the recording of the webinar here.
“Is there any advice for implementing this form of structured content in a large Knowledge Base? Any common pitfalls that should be avoided? We’re in the early stages of implementing something similar right now for a Knowledge Base with around 10000 pieces of content and it feels like a really daunting task!”
First of all, you need to have a very clear understanding of your goals both in the short-run and long-run. It’s critical to focus on your high priority goals in the first place, but it’s also important to understand how trends in your industry may affect your strategic goals. Having a big picture helps you choose a technology that will enable you to implement your strategic road map.
The correlation between the goals and technology is two-folded. On one hand, your goals determine how the information architecture should be built and what technology and tools you should use. On the other hand, the new possibilities that the technology provides may help you see your goals from a new perspective and maybe even adjust the strategy.
For example, if your immediate goals are about improving content reuse and enforcing consistency across all your content, moving to structured content sounds like a good idea. However, as you already know, once your content is semantically structured, the whole new world opens in front of you. It’s not only about content reuse any longer, it’s about an ability to deliver context-dependent content to your users, using totally new delivery channels, assembling and aggregating content automatically on-demand, and much much more.
So try to think now only about where you want to be in a few months from now, but also about where the technology enables to be in one or two years from now.
Second, make sure that the way your content is split into stand-alone pieces of information, semantically marked up, and tagged with metadata is aligned with your goals and possible contexts of your users. For example, if the context is usually very granular (a maintenance engineer searches for information how to perform scheduled maintenance for release 5.0 of component A in flavor X of machine Z) while a minimum information unit for your content is a section which includes maintenance information about all components for all flavors of the machine, you have to review your information architecture.
Third, don’t try to implement everything at once and in one go. This would be too risky. Run a pilot with a limited scope (which doesn’t look so daunting as an entire big implementation) to test your ideas and goals. You may need to do several iterations until the approaches, technologies, and tools become aligned with your goals, budget, and other constraints. There is an amazing book that I recommend to read to anyone who is going to implement a big project. It’s The Lean Startup by Eric Ries. It’s useful not only for startup founders and entrepreneurs, but to anyone who wants to start something new and innovative.
Fourth, try to see what can be automated or semi-automated. There might be a bunch of boring tasks that can be automated. For example, this includes splitting legacy content to stand-alone pieces of information, tagging content with semantic metadata, identifying potentially reusable content, and so on.
“Is there a technical and language style guide that can be referred to, for authoring content for the chatbot? Can you give some pointers towards it?”
I’m not aware of any “official” style guide for writing content for chatbots which would be similar to technical writing style guides released by companies like Microsoft or Google. Perhaps, it’s because what and how you write depends on what you want your chatbot to do. Is it a customer support chatbot? Is it a virtual assistant that helps you choose which model of the product you need? Is it a chatbot that helps your company get a new customer onboard?
Depending what you want to achieve, the type and style of the conversation will be different. However, there are some general advises, and I’ve put together some links that you may find useful:
“What is the best method to identify the failure of proper answer from chatbot? What accordingly to you is a good way to automate the learning based on failure of proper response from chatbot?”
In some online help systems, in the very end of a topic, there is a question “Was it helpful?” with an ability to click Yes or No. A chatbot could ask such a question too after answering the user’s question. The feedback provided by the user along with the conversation log can be used to track what went wrong: whether the content itself wasn’t written well, the chatbot picked up the wrong piece of information from the repository, or the user’s question wasn’t understood correctly.
But relying on the user’s feedback wouldn’t be enough. The chatbot itself should know the probability of the right answer. A customer support chatbot that we are building right now automatically calculates a special score (from 0 to 1) that indicates the level of chatbot’s confidence that the answer it gave is correct. For example, the score of 0.9 means the chatbot is almost sure that the found answer addresses the user’s question. The score of 0.6 means the chatbot has found some information that might answer the user’s question, but not sure if it’s complete or specific enough. The score is displayed to a human trainer who compares it with the actual relevance of the answer and points the chatbot to the right answer, if necessary. This allows the chatbot to learn how to search information when a similar question will be asked in the future.
“How virtual assistant (for example Microsoft Word virtual assistant) different from chatbot?”
To put it simply, chatbots tell you what to do and how to do, while virtual assistants actually do this task for you.
“Thank you for an enlivening session. Do chatbots work with desktop (PCs) too? or only on smartphones?”
Chatbots work in messengers, like Skype, Facebook Messenger, Whats App, or Slack. Most messengers of these messengers work on both desktop computers and smartphones. But it could be also a messenger embedded into your website so that your visitors can talk to the chatbot directly from the website.
In the next post in the series, I’m going to answer questions related to chatbot platforms and tools.