Keyword Extractor

About the Keyword Extractor

Counting words is easy and mostly useless. The words a document repeats are often not the words it is about — a page arguing one point will repeat the connective machinery of the argument far more often than the term at its centre, and any frequency-based extraction will hand you the machinery.

This page ranks by centrality instead: how much a term matters to what the text argues, rather than how many times it appears. A term used twice in the thesis outranks one used nine times in an aside. That is a judgement rather than a count, which is the reason this tool needs a model and the word counter on this site does not.

Three output shapes cover the uses. A ranked list is for reading. Grouping under themes is for a long document where the terms fall into distinct clusters, and the clusters themselves are often the finding. Comma-separated output is for pasting into something else.

Multi-word phrases are included by default, and turning them off changes the results substantially. Where the phrase is the unit of meaning, splitting it destroys it: "browser-based processing" is a concept, while "browser" and "processing" separately are two ordinary words that appear in millions of documents. Most of the terms worth having in a technical or commercial document are phrases of two or three words.

Stop words are excluded, along with generic filler nouns — thing, way, process, solution — unless the document is using one as a term of art, which does happen. So is anything that would describe a thousand other documents equally well, which is the real test of whether an extracted keyword is worth anything.

The document's own wording is preserved. If a page consistently says "handset", the term you get back is handset, not "mobile phone" normalised into what somebody thought it should have been called.

How to use it

1. 1Paste the article, page or batch of messages you want read.
2. 2Choose a shape: ranked to read, grouped to see the clusters, comma-separated to paste elsewhere.
3. 3Set how many terms you want. Twenty is a sensible default for a single article.
4. 4Leave multi-word phrases on unless you specifically need single words.
5. 5Compare the result against what you intended the page to be about. A mismatch is the useful finding.

Questions

How is this different from counting word frequency?

Frequency measures repetition; this measures centrality. They disagree constantly. An article about database indexing will say the word database forty times and the word index eleven times, and index is the subject. A frequency tool cannot see that difference and a model reading for sense can.

Can I use it for keyword research?

Not on its own, and it is worth being clear about the limit. This tells you what a text is about. It has no access to search volume, competition or what anyone is actually typing into a search engine. It is useful for auditing what your existing pages cover, and for finding gaps, rather than for choosing what to write next.

Why did a term I care about not appear?

Usually because the text mentions it without being about it. That is worth knowing rather than working around: if a term matters to you and does not survive extraction from your own page, the page probably does not make enough of it.

Does it work on a batch of customer messages?

Yes, and it is one of the more useful applications — paste a column of support tickets or survey answers and the recurring subjects surface quickly. Pair it with the sentiment page if you also want to know how people feel about each one.