Skip to content

Rust library providing fast language model queries in compressed space

License

Notifications You must be signed in to change notification settings

kampersanda/tongrams-rs

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

tongrams-rs: Tons of N-grams in Rust

Documentation Crates.io License: MIT

This is a Rust port of tongrams to index and query large language models in compressed space, in which the data structures are presented in the following papers:

What can do

  • Store N-gram language models with frequency counts.

  • Look up N-grams to get the frequency counts.

Features

  • Compressed language model. tongrams-rs can store large N-gram language models in very compressed space. For example, the word N-gram datasets (N=1..5) in test_data are stored in only 2.6 bytes per gram.

  • Time and memory efficiency. tongrams-rs employs Elias-Fano Trie, which cleverly encodes a trie data structure consisting of N-grams through Elias-Fano codes, enabling fast lookups in compressed space.

  • Pure Rust. tongrams-rs is written only in Rust and can be easily pluged into your Rust codes.

Input data format

The file format of N-gram counts files is the same as that used in tongrams, a modified Google format, where

  • one separate file for each distinct value of N (order) lists one gram per row,
  • each header row <number_of_grams> indicates the number of N-grams in the file,
  • tokens in a gram <gram> are sparated by a space (e.g., the same time), and
  • a gram <gram> and the count <count> are sparated by a horizontal tab.
<number_of_grams>
<gram1><TAB><count1>
<gram2><TAB><count2>
<gram3><TAB><count3>
...

For example,

61516
the // parent	1
the function is	22
the function a	4
the function to	1
the function and	1
...

Command line tools

tools provides some command line tools to enjoy this library. In the following, the example usages are presented using N-gram counts files in test_data copied from tongrams.

1. Sorting

To build the trie index, you need to sort your N-gram counts files. First, prepare unigram counts files sorted by the counts for making a resulting index smaller, as

$ cat test_data/1-grams.sorted
8761
the	3681
is	1869
a	1778
of	1672
to	1638
and	1202
...

By using the unigram file as a vocabulary, the executable sort_grams sorts a N-gram counts file.

Here, we sort an unsorted bigram counts file, as

$ cat test_data/2-grams
38900
ways than	1
may come	1
frequent causes	1
way has	1
in which	14
...

You can sort the bigram file (in a gzip format) and write test_data/2-grams.sorted with the following command:

$ cargo run --release -p tools --bin sort_grams -- -i test_data/2-grams.gz -v test_data/1-grams.sorted.gz -o test_data/2-grams.sorted
Loading the vocabulary: "test_data/1-grams.sorted.gz"
Loading the records: "test_data/2-grams.gz"
Sorting the records
Writing the index into "test_data/2-grams.sorted.gz"

The output file format can be specified with -f, and the default setting is .gz. The resulting file will be

$ cat test_data/2-grams.sorted
38900
the //	1
the function	94
the if	3
the code	126
the compiler	117
...

2. Indexing

The executable index builds a language model from (sorted) N-gram counts files, named <order>-grams.sorted.gz, and writes it into a binary file. The input file format can be specified with -f, and the default setting is .gz.

For example, the following command builds a language model from N-gram counts files (N=1..5) placed in directory test_data and writes it into index.bin.

$ cargo run --release -p tools --bin index -- -n 5 -i test_data -o index.bin
Input files: ["test_data/1-grams.sorted.gz", "test_data/2-grams.sorted.gz", "test_data/3-grams.sorted.gz", "test_data/4-grams.sorted.gz", "test_data/5-grams.sorted.gz"]
Counstructing the index...
Elapsed time: 0.190 [sec]
252550 grams are stored.
Writing the index into "index.bin"...
Index size: 659366 bytes (0.629 MiB)
Bytes per gram: 2.611 bytes

As the standard output shows, the model file takes only 2.6 bytes per gram.

3. Lookup

The executable lookup provides a demo to lookup N-grams, as follows.

$ cargo run --release -p tools --bin lookup -- -i index.bin 
Loading the index from "index.bin"...
Performing the lookup...
> take advantage
count = 8
> only 64-bit execution
count = 1
> Elias Fano
Not found
> 
Good bye!

4. Memory statistics

The executable stats shows the breakdowns of memory usages for each component.

$ cargo run --release -p tools --bin stats -- -i index.bin
Loading the index from "index.bin"...
{"arrays":[{"pointers":5927,"token_ids":55186},{"pointers":19745,"token_ids":92416},{"pointers":25853,"token_ids":107094},{"pointers":28135,"token_ids":111994}],"count_ranks":[{"count_ranks":5350},{"count_ranks":12106},{"count_ranks":13976},{"count_ranks":14582},{"count_ranks":14802}],"counts":[{"count":296},{"count":136},{"count":72},{"count":56},{"count":56}],"vocab":{"data":151560}}

Benchmark

At the directory bench, you can measure lookup times using N-gram data in test_data with the following command:

$ RUSTFLAGS="-C target-cpu=native" cargo bench
count_lookup/tongrams/EliasFanoTrieCountLm
                        time:   [3.1818 ms 3.1867 ms 3.1936 ms]

The reported time is the total elapsed time for looking up 5K random grams. The above result was actually obtained on my laptop PC (Intel i7, 16GB RAM), i.e., EliasFanoTrieCountLm can look up a gram in 0.64 micro sec on average.

Todo

  • Add fast elias-fano and pertitioned elias-fano
  • Add minimal perfect hashing
  • Add remapping
  • Support probability scores
  • Make sucds::EliasFano faster

Licensing

This library is free software provided under MIT.

About

Rust library providing fast language model queries in compressed space

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages