1 unstable release

0.1.0 Jun 1, 2024

#1067 in Text processing

Download history 208/week @ 2024-05-27 27/week @ 2024-06-03

235 downloads per month

MIT license

70KB
1.5K SLoC

vidyut-kosha

A compact Sanskrit lexicon

vidyut-kosha defines a key-value store that can compactly map tens of millions of Sanskrit words to their inflectional data. Depending on the application, storage costs can be as low as 1 byte per word. This storage efficiency comes at the cost of increased lookup time, but in practice, we have found that this increase is negligible and well worth the efficiency gains elsewhere.

For your convenience, vidyut-kosha contains helper scripts that will automatically build an interesting and comprehensive dictionary. For details, see the Usage section below.

This crate is under active development as part of the Ambuda project. If you enjoy our work and wish to contribute to it, we encourage you to join our Discord server, where you can meet other Sanskrit programmers and enthusiasts.

Overview

Sanskrit is a highly inflected language, and any reasonable Sanskrit word list will have tens of millions of words at minimum. In practice, Sanskrit programs make various compromises on this word list so that performance doesn't suffer.

One common compromise is to store irregular words as-is and store regular words just in their stem form. Then at query time, we guess at the query's underlying stem with the help of prefix and suffix tables that we curate manually.

This approach is workable, but has two main weaknesses:

  • Speed. To illustrate, the query ubhe could yield the candidates ubha, ubhA, ubhi, ubh, and ubhe, all of which we must check against the store.

  • Overgeneration. By using prefix and suffix tables, we will potentially accept invalid words like *sundarA (for sundarI) or *putrakA (for putrikA). This can be useful for some applications, but often it's a decision made by need, not by choice.

vidyut-kosha avoids these weaknesses. It stores words exactly, which avoids overgeneration. And although single-key lookup is slightly slower, it avoids the multiple lookup problem, which makes it faster for many applications.

In exchange, we must pay the following price:

  • We must construct the dictionary ahead of time. If we wish to change the dictionary, we must construct it again. Construction takes several minutes to complete.

  • We must be able to convert our values to 64-bit integers. This sounds daunting but is straightforward in practice. (If you plan to use vidyut-kosha as an inflectional dictionary, our library seamlessly manages this conversion for you already.)

Usage

Detailed notes are coming soon. For now, try running the following command:

$ make create_kosha

If you have a specific word list you want to use instead, you can use the builder API directly like so:

use vidyut_kosha::{Kosha, Builder};
use vidyut_kosha::morph::*;

let mut builder = Builder::new("output-dir").unwrap();
builder.insert("Bavati", &Pada::Tinanta(Tinanta {
    dhatu: Dhatu("BU".to_string()),
    purusha: Purusha::Prathama,
    vacana: Vacana::Eka,
    lakara: Lakara::Lat,
    pada: PadaPrayoga::Parasmaipada,
}));
builder.finish().unwrap();

let kosha = Kosha::new("output-dir");

Design

Although vidyut-kosha can support any encoding type for keys, we strongly recommend using SLP1 or WX for better space efficiency and query performance.

Our underlying data structure is a finite-state transducer (FST) as implemented in the fst crate. For details on how FSTs work, we recommend this blog post by the creator of the fst crate, particularly the section titled "Finite state machines as data structures."

Briefly, an FST generalizes the trie data structure by storing both shared prefixes and shared suffixes. The construction algorithm for FSTs will find these prefixes and suffixes automatically, and it will try reusing saved prefixes and suffixes wherever possible.

FSTs have several important constraints, two of which are most revelant here:

  1. The values in the FST must satisfy a specific algebra. Rust's fst crate satisfies this algebra by requiring that values are 64-bit integers. So to put it simply, we must be able to convert whatever information we need to store into a 64-bit integer.

  2. The keys in the FST must all be unique. But Sanskrit words are not necessarily unique, and we must to be able to store duplicate keys.

Solving the encoding constraint

For the use case of an inflectional dictionary, we want to store two kinds of information: enumerated data and text data.

Enumerated data includes categories like person, number, gender, case, and so on. Each category can take one of several possible values that we know ahead of time. We can trivially convert such data to an integer value. For example, we represent the vacana (number) of a word with one of four possible values: eka, dvi, bahu, or none if the vacana is unknown for some reason. Thus we can map eka to 1, dvi to 2, and so on.

Text data includes the stem or root of the underlying form, and we cannot feasibly enumerate it ahead of time. We encode this data through a lookup table approach: if we append all strings to a list, then the index of that string in the list is its integer representation. By using this approach, we pay the price of storing these strings the old-fashioned way. But a list of lemmas is much smaller than a list of words, so the space requirements are trivial.

Once we have mapped all of our information to integer values, we can treat our 64-bit integer as a bitfield and add values together with the appropriate shifts. For example, here's an early version of our scheme for tinantas, which uses only 32 bits:

OOLLLLppVVaadddddddddddddddddddd

O: part of speech (2 bits = 4 possible values)
L: lakara (4 bits = 16 possible values)
p: purusha (2 bits = 4 possible values)
V: vacana (2 bits = 4 possible values)
a: pada + prayoga (2 bits = 4 possible values)
d: dhatu ID (20 bits = ~1 million possible values)

One real consequence of the encoding constraint is that we can't associate words with completely arbitrary data. But if our data is structured carefully, we have plenty of room to associate each word with interesting information.

Solving the uniqueness constraint

A Sanskrit dictionary is practical only if it support duplicate keys. To solve the uniqueness constraint, we use the following workaround, which (surprisingly) works well in practice.

The fst crate treats keys as a vector of unsigned 8-bit integers, which means that each "letter" in the key can take a value in the range [0, 255]. Since the ASCII range starts at 65, we can encode extra information in the range [0, 64], which lets us mark duplicates.

For example, we might store the various forms of gacchati as follows;

  • gacchati,
  • gacchati + 0
  • gacchati + 1

And so on up to gacchati + 64 if necessary. This simple approach lets us store up to 65 duplicates of a given key.

Since even 65 duplicates is not enough in practice, we simply extended the range of bytes we append to the key:

  • gacchati,
  • gacchati + 0 0
  • gacchati + 0 1,
  • gacchati + 0 2

And so on up to gacchati + 64 64 if necessary. This approach lets us store more than 4000 duplicates per key, which is more than enough for Sanskrit.

At lookup time, we would first check for the query word -- gacchati, let's say -- then sequentially chehck gacchati + 0 0 and so on until we no longer find any results. This lookup might seem extremely expensive, but due to how FSTs are structured, it is surprisingly cheap.

Dependencies

~6–15MB
~156K SLoC