Many online interactions today are navigated with the help of a chatbot. Chatbots and conversational apps provide immediate assistance for customers and alleviate frontline support and sales staff from repetitive question-answering tasks, allowing them to focus on unique cases that require their expertise.
But current chatbot interactions are limited, guided by rules-driven workflows that require engineers to preemptively guess questions users will ask and program canned responses to those questions. The user experience with a chatbot is consequently unadaptive and static.
By the same token, if a chatbot doesn’t have a pre-programed answer to a customer’s question, a series of forms that need to be completed by the user may be presented, often culminating in a “thanks for contacting us, we’ll get back to you” message. This delays a customer’s opportunity to complete their purchase and likely sends them off to another website that can immediately provide an answer to their question and a product to solve their problem.
Deep Learning Makes Chatbots Better
It doesn’t have to be this way. At his talk at Activate Product Discovery earlier this year, Lucidworks Senior Solutions Engineer, Steven Mierop, shared a solution that empowers a chatbot with deep learning models to provide your customers with a better and smarter experience. Watch the full talk here:
In this session, Mierop explains how Smart Answers, a Lucidworks Fusion add-on, uses deep learning to understand relationships between words and extract contextual meaning. This enables your chatbot to produce relevant, opportune responses to customer questions. Smart Answers can be embedded into any chatbot framework that’s capable of making a call to an external system (including Google Dialogflow and Rasa among many others).
Smart Answers’ flexibility also allowed it to be embedded directly into query responses, and displayed next to products, images, or in the HTML.
Where Are Answers Sourced?
Smart Answers ingests data from multiple sources, including FAQ documents, user reviews, support manuals, online glossaries, product details, attribute information, and more. Question/answer pairs are created by parsing out items such as header tags and the paragraphs beneath them.
Since data comes from multiple sources, derived from multiple user types, different kinds of questions can be answered. Let’s use a customer who is interested in an exercise bike as an example. On an exercise equipment retailer’s website, he asks a chatbot an informational question like “what is a flywheel,” Smart Answers determines that the best response comes from the product manufacturer’s provided information, something like an FAQ.
For a comparative question like “is this bike really sturdy” or “what’s your experience with this bike” Smart Answers determines the answer is best sourced from the voices of other exercise bike customers, via customer reviews. And of course, although the sources of the responses vary, all of these questions can be asked and answered in one place, the retailer’s chatbot, making a simple and superior customer experience.
Different Words, Same Meaning
Natural language processing (NLP) and semantic search capabilities in Smart Answers allow customers to ask questions in an unprescribed way. Questions with different wording, but the same meaning behind the words, will produce the same response from the chatbot, even when the words are not included in the index.
Mierop used outdoor retailer and Lucidworks customer KÜHL’s data to illustrate one of Smart Answers’ uses. From a product page on KÜHL’s website, he clicks on a chatbot to ask “how can I cancel an order?”. The KÜHL chatbot responds with its cancellation policy. “That was a great response, it was accurate,” Mierop says, “but it wasn’t too impressive as any search engine worth its salt that has good relevancy tuning would most likely return that response.”
Smart Answers’ greatest payout, Mierop argues, lies in its ability to appropriately respond to a question that isn’t worded in a standard way, as none of a query’s keywords need to be mapped to synonyms to serve up a meaningful response. “How do I undo an order?” Mierop asks the chatbot. Even though “undo” is not included anywhere in the body of the response, Smart Answers understands that the latter question has identical intent to the former and so provides KÜHL’s cancellation policy. Similarly, “I need to stop my order” provides the same chatbot reply.
With deep learning in its back pocket, Smart Answers doesn’t require the upkeep of a list of rewrites and synonyms, alleviating time consuming grunt work, and missed sales and support opportunities.
Smart Answers is not only applicable for standard retail. Mierop demonstrates this point with fictitious financial services company, Midwest Mutual, who sells auto and life insurance. “What’s a trust?” Mierop asks its chatbot. Behind the scenes, Smart Answers identifies that, from its multiple data sources, this question is best answered by a glossary definition. Fusion’s NLP capabilities provide the answer in easy-to-digest conversational language.
Asking ”Why should I use a trust?” surfaces an answer from an FAQ, but the answer could also be found and extracted from, say, a blog post. When asking “How do I change a beneficiary” or “How do I alter an inheritor” Smart Answers identifies the same intent behind the two questions and responds with the same answer, whether or not the search terms exist in the response.
How to Implement Smart Answers
In order to get Smart Answers up and running, once data sources are identified, they need to be connected and then ingested into a search index. Data enrichment via a rules engine allows an engineer to program “If x and y, then z,” to do NLP, and to incorporate shopper feedback. For example, when a shopper gives a thumbs up that an answer was relevant to his question, that can be captured as a signal and used to boost appropriate responses for future customers.
With this infrastructure in place, the added benefit is that not only is Smart Answers possible, but additional product discovery and customer service solutions can be plugged in.
The ultimate purpose, of course, is to improve the shopping and support experience for your customers. Shoppers who get their questions immediately answered are happier and more self-reliant so they can progress through the sales funnel without hitting any roadblocks. “And,” Mierop says, “the best part is you don’t have to maintain that really wild web of conditional logic to create the Q and A experience.”
Learn more about the impact conversational applications with deep learning can have for your business in this blog: “Measuring the ROI of Conversational Apps.”