- sort the words in the document
- eliminate unique words (they sort together)
- merge the sorted words with the sorted dictionary and keep only the missing words
Now this blew people away when it came out
https://winworldpc.com/product/turbo-lightning/1x
it had a compressed dictionary that would fit together with the other programs you were running on a PC and spell check as you typed; there was a 640k limit for the PC but it could only use a fraction of that so as not to interfere and in the early days of the PC you couldn't actually afford to fill it out.
Someone probably could look all this up in the "unified version control of all Unix" history, but the way Doug describes it in the 1982 paper linked in the article, it sounds like the v6 spell did a "doc word at a time loop" instead of "sort all doc words at once and merge against a pre-sorted dictionary", and in fact it sounds like Johnson selected a small dictionary to facilitate that instead of "merging".
I guess you could save the working copy to disk first, then do the sort, then compare, then reload the working copy. I think the C=64 developers probably avoided that strategy because the disk interface was so damn slow.
A Bloom filter is a specific type of superimposed code.
Calvin Mooers developed random (1) superimposed coding in his Master's thesis at MIT back in the 1940s, directly influenced by Shannon's work.
Bourne's superb 1963 book "Methods of Information Handling" gives details of the mathematics.
I've no doubt Douglas knew about the broader technique, which, for example, the author of "The Large Data Base File Structure Dilemma" (1975) at http://dx.doi.org/10.1021/ci60001a005 described as "an old technique called super-imposed coding".
(1) The "random" is an important qualifier because there were superimposed codes predating Mooers, but they were not mathematically interesting or all that practically important.
The main memory bandwidth and the disk bandwidth were about the same, a little of 1MB/s.
I would have done this in multiple passes (but still used the Bloom Filters, those are cool).
https://github.com/arnoldrobbins/v10spell
https://code.google.com/archive/p/unix-spell/
The original paper is great https://www.semanticscholar.org/paper/Development-of-a-Spell...
It is hosted on his web page https://www.cs.dartmouth.edu/~doug/
https://en.wikipedia.org/wiki/Douglas_McIlroy
If you are a word nerd, you will have found obovate and there, this chart.
https://upload.wikimedia.org/wikipedia/commons/e/e8/Leaf_mor...
https://vintageapple.org/pcworld/pdf/PC_World_8711_November_... page 237 has a review (big PDF warning)
Wow, what a thing for a tech publication to suggest doing! Different times.
- They had a "dictionary" of 30000 words, and accepting a ~1/4000 rate of false positives meant that if they hashed each word to a 27-bit string (integer), they could throw away the dictionary and the problem reduces to storing a set of 30000 27-bit strings.
- Somewhat surprisingly, information theory tells us that 30000 27-bit strings can be stored using not 27 but just ~13.57 bits per word. I understand the math (it's straightforward: https://www.wolframalpha.com/input?i=log_2%282%5E27+choose+3... ) but it will probably take me a while to stop finding this counterintuitive, as 30000 is so small compared to 2^27 (which is ~134 million) that it is hard to see where the gains come from.
- To encode this 30000-sized subset of 27-bit hashes, they used hash differences, which turn out to be geometrically distributed, and a coding scheme tuned for geometrically distributed input (Golomb coding), to actually achieve ~13.6 bits per word.
I've tried to think of how one could do better, even in principle and with infinite time, along the lines of “perfect hashing” — maybe there should be a function that will take an alphabetic word, do some transformations on it, and the resulting hash will be easy to verify for being in the good set vs not. But thinking about it a bit more, the fact that we need that false-positive rate (non-dictionary words shouldn't get mapped to anything in the "good" set) requires us to use at least 27 bits for the hash. What they did seems basically theoretically optimal? Or can there exist a way to map each word to a 27-bit integer, such that the good strings are those with values less than 30000, say?
I've not decompiled it to see how it works, but it's small and fast, and works well.
https://dflund.se/~pi/cpm/files/ftp.mayn.de/pub/cpm/archive/...
Related, a contest about compressing the Wordle dictionary: http://golf.horse/wordle/
I had a lot of data, 640KB RAM, 64KB of heap, and 64KB of stack. I had hundreds of megabytes that I had to search extract data from and then combine some of them.
I experimented with data index structured into ternary trees. Conceptually it made sense, but implementation-wise the relationships and paths were still too big to keep in 64KB.
Instead of compression, I did swapping. I wrote a TSR (think service), a piece of code that would process a chunk of the data, extract the results, store it n the stack, dump the original data, make an interrupt call to the TSR, which in turn destroy the heap, and read in the next chunk from storage, return control to the program, process, combine with stack data, and continue until finished the entire process.
Originally this process took about a week for three data entry persons (think about a dozen 3" ring binders filled with tables), and an specialist combining the information. The program completed the work in just a few hours. It was amazingly "fast".
This was on a single threaded system.
[0] https://en.wikipedia.org/wiki/Terminate-and-stay-resident_pr...
UNIX spell was the 'ZX81 1K chess' of spelling, and, on home computers, we did not have a lot of spell checking going on until MS Word for Windows 3.1. Before then, in offices, secretaries did the typing with WordPerfect. They were human spell checkers for their respective managers and teams.
Meanwhile, at home, with our dot matrix printers and flickery screens, we were winging it with paper dictionaries for all of those early years of computing. I can't remember spell checking as being that important back then as everyone could spell. I was in a school of a thousand and there was only one kid that claimed to be dyslexic, a plausible excuse for not being able to spell. Maybe the 1980s was literacy's golden age with there being a clear start date for the decline in our spelling ability, that being the day UNIX spell was written.
I like to play Scrabble. Although a very different problem to spell checking, the process shares some steps with UNIX spell. Common word prefixes and suffixes are identified and bolted together in the rack or on the board with other components. Then a Scrabble dictionary is a bit like UNIX spell as it is just a big dictionary of words with no meanings provided. All that matters is whether a given word is in the book or not. It also has a few special look up tables such as the 102 two letter words.
And it is a completely different thing. In general, it is more about procedural generation and tricks then good packing. Runtime packers are used, like crinkler and kkrunchy, but actually they use a lot of RAM, like hundreds of MB, which is a bit surprising considering that the decompressed executable is in the tens of kB. But that's because they use very powerful but slow compression algorithms.
Sizecoding usually doesn't care about RAM, unless the platform requires it, the only think that matters is the size of the executable file and its data. For that 39kB Playdate game, I guess that's the same idea. The Playdate has 16MB of RAM, I bet the game took full advantage of it.
makewords sentence | lowercase | sort | unique | mismatch
I still use UNIX every day primarily for text processing.
Back in the early 1980's, I recall futzing around a bit with dictionary compression algorithms. Very interesting - but it didn't take me long to conclude that outperforming `spell` wouldn't be easy. Nor worth the effort.
This is pretty much the real thing: https://github.com/robpike/typo
(If you're crabby about it, this is similar: https://github.com/crate-ci/typos/)
Since when can we subdivide bits!?
They are trying, but only seem to go backwards. See Windows, Android, iOS, Teams for examples.
I'm frustrated when people put this kind of typography on the web. HTML can do superscript.
