Use the Blank Sheet of Paper Test to Optimize for Natural Language Processing

Posted by Evan_Hall

If you handed someone a blank sheet of paper and the only thing written on it was the page’s title, would they understand what the deed made? Would they have a clear idea of what the actual document might be about? If so, then congratulations! You precisely extended the Blank Sheet of Paper Test for sheet entitlements because your title was descriptive.

The Blank Sheet of Paper Test( BSoPT ) got any idea Ian Lurie has talked about a lot over the years, and recently on his new website. It’s a test to see if what you’ve written has meaning to someone who has never encountered your label or material before. In Ian’s terms, “Will this verse, written on a blank sheet of paper, make sense to a stranger? ” The Blank Sheet of Paper Test is about clarity without context.

But what if we’re performing the BSoPT on a machine instead of a person? Does our reflect experiment still apply? I think so. Machines can’t read–even sophisticated ones like Google and Bing. They can only guess at the meaning of our material, which makes the test peculiarly relevant.

I have an alternative version of the BSoPT, but for machines: If all a machine could see is a list of words that appear in a document and how often, could it reasonably guess what the document is about?

The Blank Sheet of Paper Test for name frequency

If you passed person a blank sheet of paper and the only thing written on it was this counter of words and frequencies, could they guess what the article is about?

An article about sharping a bayonet is a pretty good guess. The commodity I made this word frequency counter from was a how-to guide for sharping a kitchen knife.

What if the words “step” and “how” appeared in the counter? Would the person reading be more confident this article is about sharpening bayonets, or less? Could they tell if this article is about sharpening kitchen knives or pocket spears?

If we can’t get a pretty good idea of what the article is about based on which messages it consumes, then it fails the BSoPT for utterance frequency.

Can we still use word frequency for BERT?

Earlier natural language processing( NLP) approaches employed by search engines used statistical analysis of word frequency and command co-occurrence to adjudicate what a page is about. They ignored the seek and part of speech of the words in our material, basically discussing our pages like bags of words.

The tools we be applicable to optimize for that kind of NLP equated the word frequency of our content against our competitors, and told us where the gaps in word usage were. Hypothetically, if we contributed those names to our content, we would rank higher, or at least help search engines understand our content better.

Those tools still exist: Grocery Muse, SEMRush, seobility, Ryte, and others have some sort of word frequency or TD-IDF gap analysis capability. I’ve been using a free term frequency implement called Online Text Comparator, and it works pretty well. Are they still helpful now that search engines have advanced with NLP comings like BERT? I think so, but it’s not as simple as more utterances= better rankings.

BERT is a lot more sophisticated than a bag-of-words approach. BERT looks at the order of words, part of speech, and any entities represented in our content. It’s robust and can be trained to do many things including question rebutting and listed entity recognition–definitely more advanced than basic parole frequency.

However, BERT still needs to look at the words present on the page to capacity, and command frequency is a basic summary of that. Now, message orientation and part of speech trouble more. We can’t exactly sprinkle the words we is located within our spread analysis around the page.

Heighten material with word frequency implements

To help form our material unambiguous to machines, we need to make it unambiguous to customers. Abbreviating ambiguity in our writing is about choosing terms that are specific to the topic we’re writing about. If our writing gives a lot of generic verbs, pronouns, and non-thematic adjectives, then not only is our content bland, it’s hard to understand.

Consider this extreme precedent of non-specific language 😛 TAGEND

“The trick to finding the claim chef’s knife is finding a good balance of peculiarities, qualities and expenditure. It should be made from metal strong enough to keep its side for a nice quantity of meter. You should have a comfortable manage that won’t draw you tired. You don’t need to spend a lot either. The home concoct doesn’t need a figment $350 Japanese knife.”

This copy isn’t huge. It looks almost machine-generated. I can’t imagine a full essay written like this would guide the BSoPT for oath frequency.

Here’s what the word frequency table looks like with some stop oaths removed:

Now suppose we applied a word frequency tool on a few sheets that are ranking well for “how to pick a chef’s knife” and found that these parts of speech were being used somewhat often 😛 TAGEND

Entities: blade, sword, lethargy, damascus sword, santoku, Shun( symbol) Verbs: control, choppingAdjectives: excellent, hard, high-carbon

Incorporating these commands into our copy would furnish text that’s significantly better 😛 TAGEND

“The trick to finding the perfect chef’s knife is getting the right balance of aspects, excellences, and toll. The blade should be made from steel hard enough to keep a sharp-worded margin after repeated implement. You should have an ergonomic administer that you can control comfortably to prevent fatigue from providing chopping. You don’t need to spend a lot, either. The home cook doesn’t need a $350 high-carbon damascus steel santoku from Shun.”

This upgraded text will be easier for machines to classify, and better for customers to read. It’s likewise just good writing to use names relevant to your topic.

Seem toward the future of NLP

Is improving our content with the Blank Sheet of Paper Test optimizing for BERT or other NLP algorithms? No, I don’t think so. I don’t think there is a special placed of words we can add to our material to magically rank higher through manipulating BERT. I see this as a route to ensure our material is understood clearly by both consumers and machines.

I anticipate that we’re getting pretty close to the point where the idea of optimizing for NLP will be considered absurd. Maybe in 10 times, writing for customers and writing for machines will be the same thing because of how far the technology has advanced. But even then, we’ll still have to make sure our material compiles smell. And the Blank Sheet of Paper Test will still be a great region to start.

Sign up for The Moz Top 10, a semimonthly mailer modernizing you on the top ten hottest sections of SEO news, tip-off, and rad attaches uncovered by the Moz team. Think of it as your exclusive accept of stuff you don’t have time to hunt down but want to read!

Read more: