Glossary: Content and Semantics

Two passages can use different words and still mean similar things. Semantic content comparison helps you measure that relationship, including how closely text relates to a subject or audience you choose. To use the results well, you need to understand what the comparison does and then read the writing itself.

Turn text into numbers you can compare

A vector is an ordered list of numbers. For semantic comparison, those numbers represent text in a shared coordinate system. An embedding is the vector a model produces, and the word also describes the process of producing it. In OppAlerts content comparisons, each representation contains 768 numbers.

The idea is that related text gets similar representations. For example, “a place to stay” and “accommodation” have no words in common, but their meanings are related. The model and the context of the whole passage still matter. Swapping one word for a synonym doesn't guarantee a particular score.

OppAlerts uses the same model for the text and the categories so it can compare them on the same basis. Two unrelated models could each produce the same number of values, but that wouldn't make their vectors interchangeable. The comparison with keyword matching explains what numerical similarity tells you that shared words alone don't.

Choose what you want the draft to be about

A taxonomy is a named set of categories covering something like subjects, audiences, industries, products, or locations. A category is one entry in that set, with its own stored representation. When you select one, you're saying what subject or audience you want to compare your draft with.

For each comparison, you can select one category per taxonomy. The report shows where that target appears in the results. The nearest categories are the ones whose representations are numerically closest to your text, listed in order. If you get ten results, those are the ten nearest candidates. They won't necessarily all describe your draft well.

Imagine a guide that explains a product category accurately but spends most of its time on how the products are manufactured. Your buyer is trying to compare two offers and needs help understanding the specifications. The topic is right, but the reader's intent is to make a decision. The guide hasn't helped them do that yet.

An audience comparison can help you spot and investigate that difference. You still need to look at the categories that came back, the target you selected, and the passage itself to decide what to change. The taxonomy guide walks through that process.

Use the result to guide your review

Similarity measures how close the representations are. Content reports display it as a percentage. But a passage can be wrong and still be very similar to its subject. A short passage on one narrow point can also match a category more closely than a thorough article covering several issues. The score doesn't establish accuracy or quality. And because no search engine takes part in the calculation, it doesn't predict rankings.

A content brief tells the writer who the audience is, which questions to answer, what evidence is available, which category to cover, and which examples would help. That gives you a purpose to compare the draft against. Without it, you can see what the text resembles, but you can't tell whether that's what the assignment needed.

Phrase suggestions are terms associated with a selected category that don't appear in your text. Use them to look for explanations you may have missed. You don't need to insert every term. If a buyer needs to understand the warranty, adding the word “warranty” won't answer their question.

The input limit is 8,192 bytes per comparison. Each ASCII character uses one byte. Accented characters, other scripts, and some punctuation use more. Longer input is rejected. For a longer article, you can compare sections separately, as long as you remember that each result describes only the section you submitted. What you submitted matters just as much as which category you selected.