You can write an external memory spell checker with a tiny amount of RAM: something like
- 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
I saw this in BASIC in Creative Computing and got it working in on my TRS-80 Color Computer which had much less than 32k of available RAM, so that was the first thing I thought when I saw the headline.
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.
Worth noting that the article mentions this alternative as their first PoC and its drawbacks: "Because of its simplistic implementation, it was not very accurate, and also slow because of dictionary lookups on the disk."
I have a feeling the algorithm used was not the smart merge mentioned in the grandparent. That "slower" spell was in v6 Unix { https://en.wikipedia.org/wiki/Spell_(Unix) } which came out in 1975 and by 1973 there were already Winchester drives doing 885 kB/s with 25ms seeks { https://en.wikipedia.org/wiki/History_of_IBM_magnetic_disk_d... }. A 250e3 word dictionary with average word length of 10 bytes would have only taken about 2500e3/885e3 = 2.8 seconds to scan and the unique words in most documents in practice would have easily fit in RAM (as mentioned in Doug's 1982 paper). Not great, but not so bad, either. People didn't spell check on every key press for another 10 years. ;-)
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".
It's easy to compress a sorted dictionary by turning something like
a
ab
abc
to
a
1b
2c
where the prefix is the number of characters shared with the last word (and might be coded as a byte as opposed to a digit like you're thinking there. This would be a simple form of compression to code up in BASIC unlike Huffman or most LZW variants which involve bit twiddling and maintaining tree data structures... Though I do remember writing SQ and USQ [1] in BASIC)
In fact, I think this smart merge technique would have been faster (and exact, with no approximations!) than what McIlroy 1982 describes on 1975..1980 hardware in spite of not fitting in memory. The end of that article says it took about 65..77 seconds of CPU on a PDP 11/70 to check a 5500 word document.
Meanwhile, the approach of the below 3 files (which would need only tiny adaptations to 1975 C) should have taken about 4 seconds at 200 kB/sec IO time (see below). This kind of technique was quite common in the 1960s database community also. So, it's a bit weird for the Bell labs guys to not have known about it, and it shows off the C/Unix system just as well. Here is an example implementation:
delta.c:
#include <stdio.h> /* stdin->out: delta encode sorted lines */
#include <string.h>
struct { unsigned nPfx:4, nSfx:4; } __attribute__((packed)) nPS;
int main(int ac, char **av) {
char word[64], prev[64];
int nPrev = 0;
while (fgets(&word[0], sizeof word, stdin)) {
int n = strlen(word), nPfx;
int m = n < nPrev ? n : nPrev;
for (nPfx = 0; nPfx < m; nPfx++)
if (prev[nPfx] != word[nPfx]) break;
nPS.nPfx = nPfx;
nPS.nSfx = n - nPfx - 1;
fwrite(&nPS, 1, 1, stdout);
fwrite(&word[nPfx], 1, n - nPfx - 1, stdout);
nPrev = n - 1;
memcpy(prev, word, n - 1);
}
return 0;
}
words.c:
#include <stdio.h> /* Emit all words in stdin, 1 to a line */
#include <ctype.h> /*~tr -cs A-Za-z \\n|tr A-Z a-z|grep -vE .{16,} */
int main(int ac, char **av) {
char wd[15];
int n, c;
#define PUT fwrite_unlocked(wd,1,n,stdout); putchar_unlocked('\n')
while ((c = getchar_unlocked()) >= 0) { /* accum word chs */
if ((c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z')) {
if (n < sizeof wd) /* make smarter & .. */
wd[n++] = tolower(c); /*..contraction aware. */
else
fprintf(stderr, "too long: %15.15s...\n", wd);
} else if (n > 0) { /* non-word c & have data */
PUT; n = 0; /* put & reset */
}
}
if (n > 0) { PUT; } /* put any final word */
return 0;
}
notIn.c:
#include <stdio.h> /* Emit stdin words not in /tmp/dict */
#include <string.h> /* Both sources must be delta-encoded */
struct { unsigned nPfx:4, nSfx:4; } __attribute__((packed)) nPS;
int main(int ac, char **av) {
FILE *fI = stdin, *fD = fopen("/tmp/dict", "r");
char wI[16], wD[16]; /* Word buffers; then lens */
int nI=0, nD=0, oI, oD; /* Flag saying last read was Ok */
#define GET(T) /* update [wno][ID] with its next word */ \
if (o##T = fread(&nPS, 1, sizeof nPS, f##T)==sizeof nPS) { \
n##T = nPS.nPfx + nPS.nSfx; \
o##T = o##T && fread(&w##T[nPS.nPfx], 1, nPS.nSfx, f##T);}
#define PUT { fwrite(wI, 1, nI, stdout); putchar('\n'); GET(I) }
GET(I); GET(D); /* set theory "not in": ordered merge impl */
while (oI && oD) {
int c = memcmp(wI, wD, nI < nD ? nI : nD);
if (c == 0) c = nI - nD;
if (c < 0) PUT /* I<D: emit&advance I */
else if (c == 0){GET(I);GET(D);} /* I==D: advance both */
else GET(D); /* I>D: advance D */
}
while (oI) PUT /* flush tail out */
} /* (echo a;cat $SOWPODS) | delta>/tmp/dict # DE UN-abridged dict
words < Input | sort -u | delta | notIn # Extract,DE,Check */
To back up my 4.2 seconds at 200 kB/sec, you can find any SOWPODS dictionary and it compresses to about 837 KiB with that delta.c. 837/200 = 4.185.
If the rejoinder is "that SOWPODS stuff had/has licensing trouble" then no problemo - just use whatever in house dictionary they used and stemming / auto-suffix junk and use that to synthesize an exact dictionary. Then you can correct it as you go and et voila. In fact, if you wanted to be even faster than this 15..20X faster and then make accuracy-perf trade-offs, then you could probably shrink IO by generating the delta-encoded stream directly from the suffix rules.
I'd recommend staying exact, though. In this case, it seems a bad algo idea led them to be both inaccurate and slower and the writing was on the wall that hardware was getting faster. And honestly, my SOWPODS dictionary that seems 15X faster may well have better coverage than what they did which means an at-the-time apples to apples might have been 20..25x faster.
As a kind of data-compression / next optimizations side-note, the 267751 SOWPODS I compressed to 857065 bytes this way can only be Zstd -19'd down to about 588040 bytes. It's all mono-case and with simply a array-of-5-bits encoding, you could honestly get that 857065 down to (857065-267751)5+26775110 bits = 703010 bytes, less than 1.2X bigger than Zstd -19, but only 1.21X better than the more naive encoding above. So, you know, simple delta encoding works like gangbusters on a sorted spelling dictionary, has a very simple set membership algo (as above), and seems like it was at least an order of magnitude faster than what they actually did instead. I'd be a bit surprised if no one pointed this out at the time.
Actually, @probably_wrong was actually right above. It took a while, but I found Bentley 1986 Programming Pearls Column 13 describes Kernighan & Plaugher 1981 as describing the original v6 spell in more detail as indeed the smart merge technique:
So, A) mea culpa & this corrects the record and B) my measurements above suggest that @Paul_Houle's delta encoding idea alone would have given ~4X IO scan reduction on unabridged with no word stemming accuracy trade-offs. { Delta coding is, of course, combinable with word stemming in the same way as McIlroy's ideas since "notIn" is just a set operator. You just would have needed to sort -u after stemming.. maybe a stem.c to replace the two trs. } So, I still don't think you needed large Random Access Memory.
I guess you really used the fact that most words are repeated to keep the byte count in check? On the old C=64 I had it was a bit of a problem not to blow out the memory with just the text of the document once you started using it for more than a 1 or 2 page paper. Keeping a second sorted copy seems almost luxurious.
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.
I may be wrong, but from "external memory" in the description I think the idea is that each of those steps can be done on disk, not RAM. An external merge sort is a pretty standard database primitive, and the other two only require purely sequential access, so are friendly to spinning disks and the like.
could have a capacity of 80 MB or so, which is (1) much larger than RAM, and (2) mainly sequential (it could fast forward for a bit and try to find a particular block, but it was pretty slow) Contrast this to the floppy drives in the 70-140 kb range which the 8-bit micros had. Thus there was a lot of literature on external memory algorithms that would work on tapes in the early days -- though similar methods are still of interest today for "big data" as RAM is faster by a lot when you access it sequentially, you want to minimize round trips in distributed systems, etc.
(It was funny though when I talked to computer scientists around 2004 and was told to forget about external memory algorithms because 'main memory' was the thing; people realized, for instance, that Google Maps could store all the tiles in RAM in half a rack of 1U servers and if utilization was high it was much cheaper than serving the tiles off disk; Hadoop came along and was a blast from the past that itself was obsolete in just a few years)
> At this time, Bloom filter was not even called Bloom filter. In his paper, Douglas calls it a “superimposed code scheme”.
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.
I can't remember the name of the product but in the 80s there was a hardware spell checker for the IBM PC. It was a box that connected between your keyboard and your PC, and if you ever typed a string of letters that it did not recognize as a dictionary word, it would beep to let you know.
One of the things that got me intrigued by Unix was an early 1980s(ish) Byte article which walked through building a (trivial example, not the "real" one) spell checker out of a split/sort/comm pipeline, something like 7 commands? 8-bit PCs didn't have anything like that, and yet it didn't look like it needed that much sophistication...
Just finished reading the article finally (thanks!). The crux of it for me was:
- 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?
For perspective, in 1983 or so, Grammatik on CP/M ran in under 64k and did "grammar checking" (spell checking, plus a bunch of expert system rules) on an 8-bit system. (It sticks in my memory because of the time spent poking at the really interesting part: that it was so compact because it was in Forth, and there was enough of an outer interpreter in the product that with a little hex editing you could just use it as a Forth interpreter - with a very specialized set of functions preloaded :-)
in the mid 80's i ran into something similar. Fast is relative.
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".
I can remember using UNIX spell with the '-b' option, because I am British. There were only two language options, and now I want to know what the decision making was behind that, how the code catered for that and where the respective dictionaries came from. Did Australians and New Zealanders use British spelling or American?
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.
I remember spell checking my essays for high school on the commodore 64, using Paperclip 64, in 1984, Before there was ANY Microsoft windows. Spell check took a few minutes, because it read the dictionary from disk as it checked, and after that you could go thru all the words that it couldn't match.
Reading about this and similar techniques in Programming Pearls (Second Edition) by Jon Bentley left the younger me spellbound. Similar to the evolution of linkers up to mold.
How about 39kB for a video game with physics, dynamic graphics, two music tracks, sound effects, online high scores, and built-in instructions? https://news.ycombinator.com/item?id=38372936
Sizecoding is a thing. .kkrieger is a rather famous 96kB FPS game. There is even an entire demoparty called Lovebyte that is dedicated to it, the biggest category is 1k, but all demoscene events I can think of have a sizecoding competition of some kind.
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.
"In order to secure funding for Unix, Ken Thompson and Dennis Ritchie pitched Unix as a text processing system for the patents department to AT&T. Naturally, a text processing system needed a spell checker as well."
I still use UNIX every day primarily for text processing.
Ah, the Good Old Days. Back when a bit of memory cost more than most butter churns, and the ultimate data transfer technology was a sharp Express rider with his saddlebags full of punch cards...
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.
Also for reference, for those who aren't familiar with typo and don't want to read the C source code listed above (side-comment: citing the comment-free source code 'for reference' is hilarious. Thank you for the laugh :) )
> But 27-bit hash codes were too big: with 2^15 words, they needed 2^15 * 27 bits of memory, while the PDP-11 had only 2^15 * 16 bits (64kB) of RAM—compression was essential.
I'm frustrated when people put this kind of typography on the web. HTML can do superscript.
Then you should enjoy this article because nearly all the expressions are presented as proper mathematical equations bar the few places where those expressions are pseudocode
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.