Jessica Peck is the Senior Consultant for Digital Analytics for CVS by day. By night, she runs all sorts of cool side experiments to become a better SEO. She’s become quite the expert on natural language processing (NLP) and machine learning as they relate to how people are searching, and as it relates to how Google is continuing to try to provide the best search results for its end users.
If you’re curious about how you should be adjusting your strategy in light of Google’s ever-increasing intelligence, this is the podcast to catch.
The highlights:
The insights:
Jessica has been online since she was 11 years old, and started doing some site-building work in college. She ended up in an administrative position where she asked to work on the company’s website. The company didn’t have a marketing budget, so she turned to SEO to help them get the word out.
Having to figure things out for herself got her into the habit of experimentation. She puts up her side sites to run them.
“I feel like you can learn things from [side sites] that aren’t things you necessarily want to risk a real site for. I like poking at the bits of an algorithm and the bits of Google that seem to be exploitable.”
She says all this poking and prodding is what launched her interest in NLP.
“I think it’s interesting to see how machines think. I think SEO is this translation between a person, Google, and the site. A person asking a question has two layers of machine interpretation before they can actually get to the content.
SEO is this application of NLP in a weird way, and I think you can learn a lot from just how machines access language.”
“NLP doesn’t necessarily have to be a one-to-one human understanding. It’s building machine understanding to get the best results for a person,” says Jessica. “And that’s a different thing.”
Jessica says that she finds one particular aspect of NLP very interesting indeed.
“It trains people to think and act differently, and then people train the machines in kind of a loop.”
She speaks of a few of her linguist friends who have done a lot of research on how the people who create these algorithms have biases.
“They might not even be able to tell that they have [them]. That ends up moving and pushing discourse. Language evolves because machines have biases built into them that we as SEOs exploit or create language around.”
“There are a lot of good avenues to start digging in.”
She recommends beginning with the Google Assistant App and the Google Cloud NLP.
The Google Cloud NLP
“You can just paste in a paragraph and they’ll split it out. Those two things together really took me to a point where I understood what NLP did and was, which I could then go and do Python.
I think creating and building Google Assistant apps is something that teaches you a lot about how people think and communicate with search in a way that you don’t really get when you’re looking at keyword lists. Because you start looking at the construction of synonyms and entities and how questions get rooted around certain ideas and which parts of a sentence are the parts a machine is going to focus on.”
Jessica says you don’t have to have a deep coding background to dive in.
“It’s super plain language. Google’s really trying to push it. Google really wants to compete with Alexa.”
“There are quite a few practical applications. On a very basic level, seeing if you can use something like an NLP algorithm to see what information a machine is reading on your page. You get the idea: these are the ideas and the entities the machines are going to think my content is focused on.”
She says that if you know how Google or Bing is looking at a page you can tweak the page to get it closer to what you’re trying to rank for.
“Or you can use stuff like Google’s APIs and look at the knowledge graph aspect of it. And see, oh, I didn’t know green computer parts is tied to raspberry pie. I haven’t mentioned any raspberry pie, but it’s relevant to my content so maybe I should tie it more into that knowledge graph.”
In short: you take the way that Google sees your website and the content you’ve created. You get a sense of what the associated entities are. Then that influences the content you build out, ultimately helping you become more visible, drive more traffic, and increase your rankings on the keywords that matter most to your business.
“NLP at scale is a fundamentally flawed exercise a lot of the time because it all depends on the size of what you’re training it on.
If you want to train an NLP algorithm to read reviews and delete them if they’re irrelevant obscene, then you’re going to have to train it on a bunch of reviews you think are good.”
So far so good. What goes wrong?
“There are going to be times when your algorithm is going to run into stuff it hasn’t ever experienced before. This is good stuff, but your algorithm says: I don’t know what this is, it must be trash. But if you widen the net too much you get the opposite problem.”
She says the key to defeating this flaw is to keep humans involved in the process.
She gives another example.
“If you’re doing translation and you want to get a machine to the point where it can translate, and this is on a big scale, if you’re eBay and you want to translate everything into Romanian, hire some Romanians to do the translations for a while. Train your algorithms. Then keep the Romanians on as editors so you can do it on a larger scale.”
She says that because humans are fundamentally bad at being logical about language, there will always be things the machine won’t be able to catch.
Despite these flaws, there are a lot of uses for NLP that aren’t limited by these algorithms.
“You can use NLP for bulk summarization of text and bulk comparison of text. I think it’s a really interesting way to do competitor analysis if you’re trying to compete on a content level. Use NLP alongside traditional content comparison methods.”
She says there are some articles that NLP does a better job of writing than humans do.
“I know some stock sites have machine learning algorithms that just analyze the stocks and write articles that are like: this stock went down. You should buy this.
There aren’t very many creative writing majors who really want to write about the meaning of Disney’s stock decreasing.”
She also brought up Amazon, and how they seem to be using NLP to analyze product reviews.
“They’ll say ‘this review mentions the quality of the product,’ or ‘here are the reviews that mention how long it took to break.’ You can’t do that with human reviewers because it’s impossible.
This is why most big tech companies are reviewing things with NLP first, and then their second level will be human reviewers, but they rely on people who are not employees to point out where the human reviewers are needed.”
She says this is how YouTube does things too.
“Have a machine scan everything, pick out the most egregious stuff, and hope your users either don’t see the problem or see the problem and report it.”
If you’re in Massachusetts, Jessica urges you to “vote Yes on 2.” This is a ranked-choice voting initiative happening statewide.
“You can rank candidates from your best to worst. This means if you really don’t want to vote for someone but you think the other alternative is way worse, you wanna throw the Green party a vote or whatever, you can do that without worrying about wasting your vote.”
If you’re not in Massachusetts she asks you to send some money to the Homeless Black Trans Women Fund.
“It’s a fund for a small community of Black trans women living on the streets of Atlanta.”
Want to get more insights from Jessica?
