Practical AI: On Natural Language Processing in Voice Commerce
“Science indicates that babies’ brains are the best learning machines ever created.”
That quote from Dr. Patricia Kuh, co-director of the Institute for Learning & Brain Sciences at the University of Washington, has made the rounds in many parenting and life science magazines.
The time has come for it to be used in discussions on artificial intelligence.
All discussions around artificial intelligence are at some level grounded in this fundamental truth: we’re trying to fuse what we know about our own brains and what we know about machines to replicate the immense learning potential of a baby’s brain, “the best learning machine ever created.”
Our founder and CEO Amar Chokhawala was an early employee at Google where he worked on using AI to improve the Gmail user experience. He thinks about AI in a similar way:
“Humans created AI by thinking about how the human brain works. A baby’s brain is the perfect example because it is continuously learning patterns and sequences through digesting sensory input.”
So, as many of us are approaching the holiday break and will be giving or receiving voice-enabled products…
…it seems more fitting than ever to examine one particular aspect of Practical AI: the interplay between Natural Language Processing (NLP) and voice commerce.
Before we dig into how NLP works, let’s level-set.
What is Voice Commerce?
Voice Commerce is intelligent voice-based search that powers intuitive shopping experiences.
Okay, but what does that mean?
If you’ve ever asked Alexa to make a purchase for you, or if you’ve ever navigated to the search bar on your favorite retailers’ site and searched by voice instead of typing, you’ve engaged in some element of voice commerce.
How we got to the point of digitally shopping through voice is a result of more technological advancements than we can possibly cover here.
However, let’s consider the simultaneous rise of three major forces that bent consumer expectations and behaviors toward voice-based digital shopping experiences.
1. Google Hummingbird
In September 2013, Google announced that it had revamped its search engine to focus not only on the keywords searched for, but also on the implied meaning of the entire search query.
This quite literally changed the game for everybody who searches for things on the web, and it continues to set a benchmark for what searchers (consumers among them) expect from their search results—whether they’re searching on Google, in an app, or on a retailer’s site.
Regarded by SEO experts as Google’s most significant search update since 2001, this shift to “semantic search,” as it is known, means that the displayed results of a query now take into account user intent, which is an incredibly challenging concept for machines to understand.
Humans can easily assess and respond to the contextual relevance of a sentence, but up until this point search engine algorithms were primarily relying on weighting keywords and learning to determine content relevancy based on what was clicked on the most.
But the user intent underlying a search for “pizza shops,” for example, was far more likely to be something like this…
…rather than something like this…
In other words, while “pizza shops” is the subject, the users’ intent is far more likely to be about finding the closest pizza shops so they can decide which is the best one to order from.
Here’s a query I just conducted for “pizza shops”:
According to Cornell University, the original Hummingbird relied “on over 200 other ranking algorithms and techniques, which include algorithms that deal with semantic analysis and natural language processing on search queries.”
Before we explore that NLP portion, let’s dive into the two other major forces at play:
2. The Growth of eCommerce
Although the history of eCommerce dates back over 40 years, Amazon is largely credited with the industry’s massive growth over the last few years.
As Business Insider reported, Amazon alone accounted for 53% of all U.S. online sales growth in 2016. Apartment mailrooms everywhere are filled with packages from goods ordered online, and in my particular apartment complex the vast majority of those packages look something like this:
Amazon’s growth isn’t happening simply because it offers a larger variety of products than its competitors; it’s happening because the company offers an unrivaled customer experience, on-site and off.
Through reducing the various consumer friction points associated with digital retail—from delivery and returns to helping consumers wade through an endless digital aisle—Amazon has effectively made it easier for consumers to buy online than to buy in a store.
Coming back to my own apartment complex, it’s not only the mailroom that’s filled with Amazon packages. I’m now seeing AmazonFresh totes (groceries that are delivered) outside of many tenants’ doors…
…even though there are a few grocery stores within walking distance.
Similar to how “Google it” has become the unquestioned way of searching for information, purchasing items on Amazon has grown into the way for digital consumers to easily purchase certain goods.
And just as Google’s Hummingbird continues to give rise to an empowered searcher who expects search engine’s everywhere to understand their intent, Amazon’s growth, thanks to the experience they offer, continues to give rise to an empowered digital consumer who expects to easily find, purchase, and receive goods.
3. The Rise of Voice Search
Some credit the rise of voice search with the 2010 launch of Google Voice Search, but two other obvious players have also propelled voice search forward:
- In October 2011, Apple introduced a beta version of Siri in the iPhone 4S
- In November 2014, Amazon released Alexa.
Note: Even Facebook seems to be building out a “Siri-like” voice assistant.
All of this points to how voice search adoption is skyrocketing. According to AdWeek, the U.S. will have 67 million voice-assisted devices in use by 2019. And ComScore predicts that by 2020, 50% of all searches will be voice searches.
With such widespread consumer adoption rates, and all of the big players investing heavily in continuing the trend and capitalizing on it (Amazon alone has 5,000 employees working on Alexa), it’s easy to see why many industry analysts consider voice-powered digital shopping the next frontier of ecommerce.
Unfortunately, the on-site search experience offered by most retailers is woefully behind.
Customers that are accustomed to Google Voice Search, Siri or Alexa expect retailers to offer an intelligent voice commerce experience, one that incorporates elements from every aspect we’ve covered so far.
Enter Natural Language Processing
Because of the immense and rising use case for retailers, NLP in voice commerce is arguably the most practical of all Practical AI examples.
But to understand it, we must circle back to Hummingbird and the way the majority of us still search today: typing.
Let’s level-set again: What is natural language processing?
NLP technologies allow for accurate, automated understandings of text and speech.
Their overarching goal is for machines to understand the natural language (typing or speaking) of humans. And because we’re in the realm of artificial intelligence here, the algorithms aren’t static; they’re continuously learning and improving.
As you can imagine, trying to get a machine to understand natural human language is quite the challenge.
To begin to understand a search query, for example, the machine must parse the parts of speech within a query. It’s like back in Grammar 101, except automated. Various parsing systems can label words with tags based on parts of speech (.n = noun, for example).
There are variety of ways this can happen, but here are two automated parsed examples of “red socks that are less than twenty dollars.”
It doesn’t end there. Us humans still misconstrue each other’s sentences, so it’s important for NLP to be able to figure out the multiple ways a sentence could be understood and then score which is the most likely given the context signals of the search (e.g. whether the search took place on Toms.com or inside an academic journal’s app).
Here’s one example from SyntaxNet, an open-source neural network framework:
Starting with “Red”
We’ll explore color more in-depth in a future Practical AI post, but for our purposes here let’s consider the color red for a moment.
To truly understand all that can be encompassed by “red,” our NLP algorithms at Reflektion go far beyond lexicon and into cognitive semantics. They use topic modeling to identify that “red” is referring to a color, and from there they map this baseline understanding to a comprehensive color hierarchy that includes every known hue of red.
Why is this so important?
For starters, because red can refer to a seemingly infinite array of reds. Let’s say an internet retailer sells red socks, but in all of the product descriptions and metadata those socks are referred to as “Scarlet.”
A search for “red socks” simply will not be able to find and then display those scarlet socks unless the algorithm understands red both as a specific color and as part of a vast neighborhood of colors—the algorithm must be capable of mapping the visual distance from one hue of red to another.
Moving to “Socks”
Similar to finding color synonyms, NLP can also understand product synonyms in order to build out a knowledge base around particular keywords. This allows a search for “socks” to map to every possible query for socks, and through parsing the sentence it can determine both what is typically meant by “red socks” and what a users’ intent typically is when they search this query.
Such knowledge can only be built from training data. A baby’s brain is processing and learning from every sensory detail; an algorithm is processing and learning from every detail it’s being told to learn from. In retail, this can include product catalogs and sources such as Google News which can be used to feed the algorithm all of the world’s articles so it can begin to develop a sophisticated understanding of how words are strung together in various contexts.
Underlying all of this, of course, is the development of a model capable of feeding the machine so that it can understand and make real-time, accurate predictions.
By leveraging NLP that is mapped to such massive datasets, digital retailers are essentially augmenting and optimizing their existing product attributes so that, for example, a search for “red socks” may display the only related product they have: a pair of “scarlet stockings.”
And finishing with “less than twenty dollars”
Understanding price-based product searches demands a deep semantic understanding because operator words that are part of product search queries, such as “under,” (as in the example above from our client O’Neill) can have various syntactical meanings.
NLP-supported operator words can include a few of the following:
- Less than
- From _ to _
- Between _ and _
Here’s a glimpse into how these processes would coalesce for a similar query, let’s say “red dresses under $100”:
Additionally, price adjectives are critical for context. This includes a few of the following:
- Cheap (including cheapest and cheaper than)
Here’s an example of “cheapest” from TOMS, another client of ours:
And then there’s “on sale,” which a retailer’s NLP-powered site should be able to understand and map to related terms, such as “discount.”
Practical AI… back to the roots
Even in the arena of voice commerce, language must go back to its digitally typed roots.
When you speak to Siri or Alexa or to a retailer’s site, you are sending data to a server that analyzes your speech and translates it into text so it can work through a few of the processes we’ve addressed here.
As voice continues to constitute larger and larger shares of all searched queries, retailers will need to adapt or they’ll quickly be viewed as behind-the-times and uncaring of customer expectations.
In a Forrester report titled, Voice Search Will Change Customer Discovery Forever, Collin Colburn implores digital sellers to ask a fundamental question to determine the urgency with which they should make the move to incorporating voice search:
“Are my target customers using voice search?”
To answer this question, Colburn suggests using Google Search Console for phrases such as “Ok, Google…” as well as assessing the longest of the long-tail queries—spoken questions tend to be far longer than typed questions.
To Colburn’s advice, I’d add factoring in a variety of demographic information. For example, a study covered in June 2017 by Search Engine Land found that 43% of millennials made a voice-device purchase in the past year.
All signs point to this number skyrocketing. Are retailers ready? Not if they’re still forcing voice-ready consumers to type their complete search, click submit, and hope.
Then you might want to know our exciting news:
We were named a top 100 AI company of 2018 by CB Insights. And get this: based on the results from CB Insights’ National Science Foundation-backed algorithm, we received the 3rd highest score of all companies on the list.
Read it here:CB Insights AI 100 Reflektion Practical AI.compressed