Spellbook is a Rust spellchecking library compatible with the Hunspell dictionary format.
fn main() {
let aff = std::fs::read_to_string("en_US.aff").unwrap();
let dic = std::fs::read_to_string("en_US.dic").unwrap();
let dict = spellbook::Dictionary::new(&aff, &dic).unwrap();
let word = std::env::args().nth(1).expect("expected a word to check");
if dict.check(&word) {
println!("{word:?} is in the dictionary.");
} else {
let mut suggestions = Vec::new();
dict.suggest(&word, &mut suggestions);
eprintln!("{word:?} is NOT in the dictionary. Did you mean {suggestions:?}?");
std::process::exit(1);
}
}
Spellbook is no_std
and only requires hashbrown
as a dependency. (Note that ahash
is included by default, see the feature flags section below.) This may change in the future for performance tweaks like small-string optimizations and maybe memchr
but the hope is to keep this library as lightweight as possible: new dependencies must considerably improve performance or correctness.
Spellbook is a work in progress and might see breaking changes to any part of the API as well as updates to the MSRV and dependencies.
Currently the check
API works well for en_US
- a relatively simple dictionary - though it should work reasonably well for most other dictionaries. Some dictionaries which use complex compounding directives may work less well.
The suggest
API was added in v0.2.0 and should behave the same as Nuspell's suggest
. (Meaning that phonetic suggestions are not implemented.)
Spellbook should be considered to be in alpha. Almost all of the Hunspell test corpus tested by Nuspell is passing.
The only feature flag currently is default-hasher
which pulls in ahash
and is enabled by default similar to the equivalent flag from hashbrown
.
A non-cryptographic hash significantly improves the time it takes to initialize a dictionary and check and suggest words. Denial-of-service attacks are not usually relevant for this use-case since you would usually not take dictionary files as arbitrary inputs, so a non-cryptographic hash is probably ok. (I am not a cryptologist.) Note that Hashbrown v0.15 and above use foldhash
instead of aHash. In my runs of the Spellbook benchmarks foldhash
doesn't make a perceptible difference.
You can easily drop this default feature:
[dependencies]
spellbook = { version = "1.0", default-features = false }
and specify a hasher of your choosing instead:
use std::hash::BuildHasherDefault;
type Dictionary = spellbook::Dictionary<BuildHasherDefault<ahash::AHasher>>;
For a more in depth overview, check out @zverok
's blog series Rebuilding the spellchecker.
Hunspell dictionaries are split into two files: <lang>.dic
and <lang>.aff
.
The .dic
file has a listing of stems and flags associated with that stem. For example en_US.dic
contains the word adventure/DRSMZG
meaning that "adventure" is a stem in the dictionary with flags D
, R
, S
, M
, Z
and G
.
The .aff
file contains a bunch of rules to use when determining if a word is correct or figuring out which words to suggest. The most intuitive of these are prefixes and suffixes. en_US
contains suffixes like R
and G
:
SFX R Y 4
SFX R 0 r e
SFX R y ier [^aeiou]y
SFX R 0 er [aeiou]y
SFX R 0 er [^ey]
SFX G Y 2
SFX G e ing e
SFX G 0 ing [^e]
Since "adventure" has these flags, these suffixes can be applied. The rules are structured as tables that define the flag (like R
), what to strip from the end of the word (0
for nothing), what to add to the end (er
for example) and under what condition the suffix applies (matches [^aeiou]y
meaning not 'a' 'e' 'i' 'o' 'u' and then 'y' for example). When checking a word like "adventurer" you find any suffixes where the "add" portion of the suffix matches the ending of the word and check if the condition applies. The first clause of R
applies since the "adventure" ends in 'e', and we add a 'r' to the end. When checking this happens in reverse. Starting with a word like "adventurer" we strip the 'r' and check the condition. Similarly with G
, the first clause matches "adventuring" because "adventure" ends with 'e' and we add an "ing".
Hunspell dictionaries use these prefixing and suffixing rules to compress the dictionary. Without prefixes and suffixes we'd need a big set of every possible conjugation of every word in the dictionary. That might be possible with the gigabytes of RAM we have today but it certainly isn't efficient.
Another way Hunspell dictionaries "compress" words like this is compounding. For example with the COMPOUNDRULE directive:
# compound rules:
# 1. [0-9]*1[0-9]th (10th, 11th, 12th, 56714th, etc.)
# 2. [0-9]*[02-9](1st|2nd|3rd|[4-9]th) (21st, 22nd, 123rd, 1234th, etc.)
COMPOUNDRULE 2
COMPOUNDRULE n*1t
COMPOUNDRULE n*mp
en_US.dic
has words for digits like 0/nm
, 0th/pt
, 1/n1
, 1st/p
, etc. The COMPOUNDRULE directive describes a regex-like pattern using flags and *
(zero-or-more) and ?
(zero-or-one) modifiers. For example the first compound rule in the table n*1t
allows a word like "10th": it matches the n
flag zero times and then "1" (the stem of the 1
flag in the .dic
file) and "0th". The n*
modifier at the front allows adding any number of any other digit, so this rule also allows words like "110th" or "10000th".
- An overview of internals
- Comparisons to other spellcheckers
- Development and contributing notes
@zverok
's blog series on rebuilding Hunspell was an invaluable resource during early prototypes. The oldspylls
-like prototype can be found on thespylls
branch.- Ultimately Nuspell's codebase became the reference for Spellbook though as C++ idioms mesh better with Rust than Python's. Nuspell's code is in great shape and is much more readable than Hunspell so for now Spellbook is essentially a Rust rewrite of Nuspell (though we may diverge in the future).
- The parser for
.dic
and.aff
files is loosely based on ZSpell.