Presented at Activate Product Discovery 2021. Today’s AI chatbots can do it all, from helping a customer to find the perfect apparel for right activity to inquiring a deal that you can get while ordering a pizza. In this session, let us demonstrate how Fusion Smart Answers could help your customers for effective product discovery.
Steven Mierop, Lucidworks Senior Sales Engineer
Hi, welcome to Real-Time Connections Using Smart Answers. My name is Steven Mierop, I’m a senior solutions engineer here at Lucidworks. Today we’re gonna be talking about intelligent question and answer and solutions, and how they can help improve the product discovery experience, as well as where you can position applications like this within different contexts and within different positions of your sales funnel.
We’ll talk about semantic search, what it means, why it’s important, and also how you can use it to power chatbots. And then finally, we’ll take a look at a couple of different demos and really tie everything together with a few live examples.
So, when you think about having questions answered online, you might immediately think about a chatbot. And traditionally, chatbots haven’t really been that performing. One of the reasons for that is they’re typically built with these really intricate static rules driven workflows, where you have to understand the potential questions or intents ahead of time and then craft a response to accommodate those. Those tend to be static in nature, whereas you really can’t get the most value and really can’t pull data from a wide knowledge phase. You’re kind of locked into what you have, and then you have to just maintain that manually over time.
Another thing you might run into when interfacing with the chatbot is you ask the question and then instead of getting an answer back, you get a series of these input forms, which ultimately lead to a message saying, thanks for your question,we’ll have someone get back to you. So really not a great experience, right.
And our solution to help solve this problem is an add-on to Fusion called Smart Answers. Now with smart answers, we can help shoppers by allowing them to ask questions in a much more natural way and then get relevant,immediate, contextual answers back. We do this by using deep learning models to help understand the relationship between words.
So if you and I were both looking for a similar answer, but we asked the question in two different ways using maybe different words, words that might even be outside of our index entirely, smart answers can help understand that relationship and then help deliver an acceptable, accurate answer back.
The benefit of using a technology like this is, you can allow shoppers to be much more self-sufficient, you can also pull more information from different data sources within your knowledge base. And the best part is you don’t have to maintain that really wild web of conditional logic to create that Q and A experience.
So what types of questions do we tend to ask online? Recently I was looking for a stationary bike and I’m not a cyclist, so I had a lot of questions around what type of components come with the bike.
For example, what type of resistance is included? What level of resistance? What is a flywheel? So, a lot of more product informational type questions. But then I also had kind of these comparative questions where I didn’t necessarily want the manufacturer to give me an answer. I want him to hear it from previous shoppers or existing customers. And the questions I was really looking for were along the lines of, what’s the difference between magnetic resistance versus friction resistance. Is the bike sturdy? What’s your experience like on it?
So, there’s a lot of different data sources and a lot of different types of questions that a shopper might ask. Now, if you’re gonna build an application like this, where would you find question and answer pairs that would eventually be put inside of our smart answers application?
And it turns out you can actually find these pairs in a lot of different places. You might find existing product reviews, you might have curated frequently asked questions, or even community driven solutions. Another really cool place to look at our support documents and online manuals or glossary lists. When you ingest this type of data through the fusion platform, you can parse out header tags, which usually will reflect a question, and then paragraphs beneath it, which might reflect the body of an answer.
So, you can pull out sometimes multiple question and answer pairs from these online pieces of documentation, maybe manuals or support documents. And then finally, you can even craft Q and A pairs from product details and attribute information.
So if you wanted to build a smart answer solution, let’s walk through some of the different components you would need to make it happen. Once you identify the different types of data, we need to connect to those data sources and then ingest them into a search index or some type of maybe sidecar collection that’s doing a nearest neighbor type of lookup. Now along the way, we still wanna do data enrichment,so we’ll need a rules engine to handle scenarios where we could say if X and Y, then Z. We also need natural language processing capabilities on top of smart answers. So things like entity extraction, parts of speech tagging,as well as a way to incorporate shopper feedback. So every time a shopper might send a positive signal, maybe a thumbs up, that the question was relevant and answered their question, we can capture that as a signal and then use signals to boost search results and make it perform better over time.
So the types of solutions we can build with this, of course are smart answers, which we’re looking at today. But we can take the same model and build other types of product discovery solutions and customer service applications. Ultimately, our goal is to make shoppers more self-reliant so that they can self-serve and answer their own questions while staying in that same context or staying within a particular sales funnel. Because the last thing we wanna do is have a shopper, ask a lot of different questions that they can’t find on the website. And then they have to bounce around to different locations where they might get distracted or potentially see an ad for another product and then completely step away from our sales funnel entirely. So, this is gonna be another way that we provide a higher level of customer care. And we’re gonna do that through more intelligent question and answering.
So what kind of experiences can you build with a technology like this? There’s a few different approaches you could take. You could use this to power a virtual assistant, where you might be verbally asking a question and then getting an answer back. You might incorporate this into a chatbot, and you can use a lot of different frameworks too. In this case, we’re using Google dialogue flow, but there’s nothing that prevents you from using a browser or some open source framework. As long as you can make a call to an external system which most modern chatbot frameworks have, you can build a solution
powered by smart answers. And then finally, you may not even need an interface like that. You might just wanna embed smart answers directly in the response of a query where you could display that right next to your product,images or within the HTML somewhere.So we have a lot of flexibility in how you can use it.
Okay, so why don’t I go ahead and show you a couple of demonstrations and you’ll get a feel for where smart answers can help. So for this first demo, we are going to walk down a product discovery path. And we’ll be using one of our existing customers data that they were generous enough to let us use. And this particular customer is Kuhl, they specialize in outdoor apparel, hiking, fly fishing and that sort.
So, I’m gonna do a search for shorts. And what you’ll see is that, we’re getting a mix of products back for all shorts, which is great, but we have men’s, we have women’s, we have different colors, different styles. Fusion hasn’t really learned what my preferences are just yet. So why don’t I add a little more detail here, I’m looking for men’s shorts. And Fusion went ahead and identified that that was a known entity that mapped it to a facet and went ahead and filtered it for me.
So now I’m getting men’s shorts coming back. This is good, and let’s say that, you know what? Maybe I wanna stand out a little bit. So I’m gonna search for red men’s shorts. And there we go. We’re getting really good precision exactly what I’m looking for. And I’ll give you a little sneak peek behind the scenes. So, what fusion is doing is it’s identifying all of these different entities within my query and it’s pulling out that, hey, red is actually a part of a product color family, men’s is related to a gender property type, or now product type is related to shorts. So you can go ahead and filter this for me and help boost different products that match what my preferences are and what I’m looking for.
So let’s say I land on these Shift Amphibia shorts, and I’m somebody who tends to get buyer’s remorse. And before I make this purchase, I wanna know that I have a good exit and I have an out clause if I need to return this. So to make me feel more confident in that purchasing decision, I might interface with a chatbot that’s being powered through smart answers.
And let’s just ask some common questions here. Like, how can I cancel an order? And we’ll play this first one so you can get a feel.
[Woman] You may cancel within 60 minutes of submitting by returning to your order history page on cool.com and clicking cancel. Need help, give us a call at 8882183181 from 6:00 AM to 8:00 PM mountain standard time.
Steven Mierop: That was a great response, right? it was accurate. But it wasn’t probably too impressive. Any search engine worth its salt that has good relevancy tuning would most likely return that. Where smart answers really helps and where its sweet spot is, is if it answers a question similar to the one I asked came in, but maybe we didn’t anticipate it, and we may not even have a synonym to map out the different words.
So let’s say instead of cancel, I say how do I undo an order? [speaker types on screen and hits enter]
You’ll notice we got that same answer back. And again, in this case, didn’t really expect that one coming in, undo wasn’t necessarily part of that the body of that answer, but it’s still identified and it returned it back to me. Or maybe I asked something entirely different. Like, I need to stop my order. And again, smart answers found that relationship, and then pulled back an accurate answer that would satisfy my question.
So, that was really helpful, and that takes a burden off of our merchandisers to have to maintain a list of rewrites and different types of synonyms to create an experience like this. Another thing kind of like in our stationary bike example, you might have questions related to attributes of a particular product. And you’ll see here we have style ID 1105. Well, you can also ask questions related to product information. You might say, tell me about 1105.
[Woman] Style idea 1105, has product named Shift Amphibia short, intended for men’s and cost $70.
Steven Mierop: So in this example, we can start extracting and pulling back relevant information that might give me more information to complete the purchase.
So let’s say for example that, maybe you’re not selling online apparel, maybe your commerce experience is more aligned with financial services, for example. So in this case, let’s take a look at a company called Midwest Mutual.
This is a fictitious insurance company that sells home insurance, auto life insurance. And as a shopper, maybe someone new to this company, I wanna ask some questions just to educate myself about their product line and about some of the terminology inside of this particular domain. So maybe ask you a question, what’s a trust?
[Man] Wills and trusts are legal documents that essentially spell out who gets what after you’re gone.
And we will listen to the whole thing. I’ll ask it again so we can see it one more time. What’s a trust?
And what happened behind the scenes here, is that we have multiple data sources just like a few slides back we looked that, where we have glossary terms, frequently asked questions, and what we can do with Fusion is identify that this incoming question is more of a definition based question, kind of like a factoid question, and we could promote or boost glossary definitions to come up ahead of frequently asked questions.So we can kind of layer in these different types of natural language processing capabilities to really build an intuitive. And engaging Q and A solution.
Now, what if I said, “why should I use a trust?” And in this case, I won’t receive a glossary definition, I might receive an actual frequently asked question or maybe something extracted from a blog post that could answer that question more accurately.
Just like in our synonym like example with apparel, we can do the same with services if we wanted to. So if I asked, “How do I change my beneficiary?” We will receive an answer back. We’ll play this first one.
[Man] To change the beneficiary. You will need to complete and sign a change of beneficiary form. Call your Midwest Mutual representative or contact Midwest Mutual directly at 18773949524, 7:00 AM to 6:00 PM Central time, Monday through Friday.
Or if someone new came along, and maybe didn’t know the vernacular very well, and didn’t use the words change and beneficiary, instead they used the words alter and inheritor, and ran that same search, Fusion would still identify that, hey, it looks like you’re really looking for this beneficiary question. And it returned that back to me to solve the problem.
So, I hope that was clear in terms of how it works and some of the different capabilities that you’re able to do, you’re able to use here. So, thanks a lot for joining and stick around, we’ll have a live Q and A following this presentation. Thank you.