Skip to content

Latest commit

 

History

History
71 lines (43 loc) · 1.91 KB

README.md

File metadata and controls

71 lines (43 loc) · 1.91 KB
Raw

sqlite3-ngram

ngram is a SQLite3 FTS5 n-gram tokenizer, it tokenize the input text in computational linguistics level.

For the input text Hello 新 世界:

  • ngram = 1

    Hello, 新, 世, 界

  • ngram = 2

    Hello, 新, 新世, 世界

  • ngram = 3

    Hello, 新, 新世, 新世界

The tokenization is based on UTF-8 character and character category boundary.

The ngram currently support is in range [1, 4], larger ngram can be supported but it's usually unnecessary.

This tokenizer extension can be used as a fallback(generic) tokenizer for FTS purpose.

Build

# Tested under podman, docker should also be ok.
container/build.sh

Usage

-- First load the ngram extension
.load build/libngram.so
-- By default N = 2, valid N is in range [1, 4]
CREATE VIRTUAL TABLE t1 USING fts5(x, tokenize = 'ngram');
CREATE VIRTUAL TABLE t1 USING fts5(x, tokenize = 'ngram gram N');

-- Or check sql/load-ext.sql for example usage
-- sqlite3 < sql/load-ext.sql

Advance usage

You can integrate this tokenizer with the SQLite3 official porter tokenizer:

.load build/libngram.so
CREATE VIRTUAL TABLE t1 USING fts5(x, tokenize = 'porter ngram gram N');

In such case, if you tokenized the word direct. directed, directing, direction, directly... all can be coalesced into direct and thus hit a match.

Limitation

Currently only the UTF-8 string is supported for tokenization, usually not a big concern though.

Credits

This project was inspired from the following projects:

wangfenjin/simple - 支持中文（简体和繁体）和拼音的 SQLite fts5 扩展

TODO

  • Implement ngram_highlight() function
  • Add more test cases
  • Enable build & test CI