Somebody counted 2,329 portmanteaus

A writer named Guy Beringer published an essay in Hunter's Weekly in 1895 arguing that the English should stop eating enormous Sunday dinners. Halfway through, he does the thing this page is about:

"The word Brunch is a corruption of breakfast and lunch, and the meal Brunch is one which combines the tea or coffee, marmalade and kindred features of the former institution with the more solid attributes of the latter." — Guy Beringer, Brunch: a plea, 1895

A corruption. He had no theory, he had a joke about marmalade, and the word outlived everything else he wrote.

The technique got its name from a different joke. In Through the Looking-Glass (1871), Humpty Dumpty explains to Alice that slithy means lithe and slimy: "You see it's like a portmanteau — there are two meanings packed up into one word." A portmanteau was a suitcase that opened into two hinged compartments, so Carroll was describing luggage. The word for the technique is itself a blend of metaphor and grammar.

For most of the following century, the serious position was that this could not be taught.

The experts said it was unteachable

Two of them said so in print, and both are quoted in the paper that eventually proved them wrong:

"in blending, the blender is apparently free to take as much or as little from either base as is felt to be necessary or desirable. […] Exactly what the restrictions are, however, is far from clear." — Laurie Bauer, English Word-Formation, 1983

"we find no discernible relationship between phonology […] and a viable blend. […] This fact helps to make blends one of the most unpredictable categories of word-formation." — Garland Cannon, 1986

That is a reasonable position held by capable people, and it lasted because nobody had counted. The classic studies in the field worked from 314 blends (Pound, 1914) and 132 (Cannon, 1986), which is enough to have opinions about and not enough to settle anything.

Then somebody counted

Stefan Gries assembled 2,329 intentional blend formations and ran the numbers (Quantitative corpus data on blend formation, in Cross-Disciplinary Perspectives on Lexical Blending, 2012). Blends turn out to be unpredictable the way weather is unpredictable, which is to say not very, once you stop looking for a rule and start looking for a distribution.

Three findings, all at p < 0.001 across 1,921 annotated blends.

The first word is shorter. Median first source word: 2 syllables, 5 phonemes, 6 letters. Median second: 3 syllables, 7 phonemes, 7 letters. Compared blend by blend rather than in aggregate, the second word runs longer by about half a syllable and one letter.

The first word is commoner. Measured against an 800,000-story newswire corpus, the first source word is significantly more frequent than the second.

Neither of those is true of accidents. Gries ran the same tests on 186 overheard slips of the tongue, where two words collide by mistake. No length difference (p > 0.8). No frequency difference (p > 0.38). The asymmetry only shows up when somebody is choosing, which makes it a description of taste.

So the rule of thumb has a number behind it now. Familiar word first, longer word second. Your reader recognizes the common short word instantly and spends their attention decoding the second one, which is the half that carries the meaning.

Worth noticing that brunch gets this backwards. Breakfast is longer than lunch and leads anyway. That exception is why Gries titled his earlier paper Shouldn't it be breakfunch? (Linguistics 42, 2004). The textbook example of the technique is a mild exception to it, and the exception is more instructive than the rule.

Why breakfunch lost

Because two forces pull against each other, and the good blends sit where they balance.

Recognizability wants to keep as much of both words as possible. By that measure alone breakfunch beats brunch outright: more of breakfast survives.

Similarity wants the blend to still look and sound like a single word rather than a collision. And this is where recognizability, pushed to its limit, becomes absurd. Gries's own demonstration:

Chevrolet + Cadillac → Chevrolecadillac

Every letter of both source words is present. Nobody has ever coined that, and nobody ever will, because the joke has been explained. As the paper puts it, such blends "are not fun anymore." Perfect recognizability, zero charm.

Somewhere between Chevrolecadillac and unintelligible mush is the band your name has to land in. Machines can measure the distance to each end of it. What they cannot tell you is where in the band your particular name should sit, which depends on how much your audience already knows about you.

The instrument

Below is every way to splice two words together, ranked by Gries's own measure: the average edit distance from the blend back to each source word. Lower means more of both words survived.

Seam finder

Every splice of the first word's front onto the second word's back, ranked by how much of both words survives.

First word+Second word

1. brunchASED 4.5clean cut
2. breunchASED 4.5clean cut
3. breaunchASED 4.5clean cut
4. breakunchASED 4.5clean cut
5. brechASED 4.5clean cut
6. breachASED 4.5clean cut

Ordered the unusual way. Across 1,921 intentional blends the first source word is the shorter one. Here it is the longer one, which puts this pair in the minority, where brunch also sits.

ASED is Gries's average string-edit distance from the blend to each source word: lower means both words stay recognizable. It reads letters, not sounds, because a name you just invented has no pronunciation entry to look up. And it measures only the half a machine can measure — Chevrolecadillac scores well and is still dead on arrival.

Try channel and tunnel, which is the worked example in the paper: chunnel scores 1.5, and the last four letters could have come from either word. Then try chevrolet and cadillac and watch the top of the ranking fill up with names no human would use. The ranking is honest about what it measures, which is one half of the problem.

The seam is the whole craft

Two blends can score identically and land completely differently, and the difference is usually the join.

Chunnel fuses. Its last four letters belong to channel and tunnel at the same time, so there is no visible seam and the word behaves like it always existed. Brunch does not fuse at all. It is a clean cut: br from one word, unch from the other, sharing nothing. It survives because both fragments are already comfortable English shapes.

The failure mode is the third case, where the seam is neither fused nor clean. You get a consonant cluster no English word would tolerate, and the name has to be spelled out every time it is spoken. That is the thing to listen for, and it is the reason to say a candidate blend aloud to somebody before you register anything.

What to do with all this

Put the short familiar word first. You now have 1,921 blends' worth of company.

Aim to lose material, not to keep it. The instinct to preserve both words is the one that produces Chevrolecadillac. A blend is a compression, and the compression is the point.

Check the seam by ear, not by eye. Fused joins read as one word. Clean cuts work when both fragments are ordinary English. Anything that makes a cluster your mouth argues with will cost you every phone call you ever make.

Notice when you are naming the category rather than the company. Blending two industry words gets you a name that describes your sector accurately and is indistinguishable from four competitors who ran the same operation on the same two words.

The technique is 131 years old on the record and probably much older in the pub. What changed recently is that we stopped guessing about it.