13 releases (6 stable)
1.1.5 | Nov 5, 2022 |
---|---|
1.1.4 | Sep 14, 2022 |
1.0.0 | Dec 22, 2021 |
0.6.0 | Dec 22, 2021 |
0.1.1 | Dec 27, 2017 |
#164 in Text processing
108 downloads per month
Used in 2 crates
320KB
1K
SLoC
wordcut-engine
Word segmentation library in Rust
Example
use wordcut_engine::load_dict;
use wordcut_engine::Wordcut;
use std::path::Path;
fn main() {
let dict_path = Path::new(concat!(
env!("CARGO_MANIFEST_DIR"),
"/dict.txt"
));
let dict = load_dict(dict_path).unwrap();
let wordcut = Wordcut::new(dict);
println!("{}", wordcut.put_delimiters("หมากินไก่", "|"));
}
Algorithm
wordcut-engine has three steps:
- Identifying clusters, which are substrings that must not be split
- Identifying edges of split directed acyclic graph (split-DAG); The program does not add edges that break any cluster to the graph.
- Tokenizing a string by finding the shortest path in the split-DAG
Identifying clusters
Identifying clusters identify which substrings that must not be split.
- Wrapping regular expressions with parentheses
For example,
[ก-ฮ]็
[ก-ฮ][่-๋]
[ก-ฮ][่-๋][ะาำ]
The above rules are wrapped with parentheses as shown below:
([ก-ฮ]็)
([ก-ฮ][่-๋])
([ก-ฮ][่-๋][ะาำ])
- Joining regular expressions with vertical bars (|)
for example,
([ก-ฮ]็)|([ก-ฮ][่-๋])|([ก-ฮ][่-๋][ะาำ])
-
Building a DFA from the joined regular expression using regex-automata
-
Creating a directed acyclic graph (DAG) by adding edges using the DFA
-
Identifying clusters following a shortest path of a DAG from step above
Note: wordcut-engine does not allow a context sensitive rule, since it hurts the performance too much. Moreover, instead of longest matching, we use a DAG, and its shortest path to contraint cluster boundary by another cluster, therefore newmm-style context sensitive rules are not required.
Identifying split-DAG edges
In contrary to identifying clusters, identifying split-DAG edges identify what must be split. Split-DAG edge makers, wordcut-engine has three types of split-DAG edge maker, that are:
- Dictionary-based maker
- Rule-based maker
- Default maker (Unk edge builder)
The dictionary-based maker traverses a prefix tree, which is particularly a trie in wordcut-engine and create an edge that matched word in the prefix tree. Rule-based maker uses regex-automata's Regex matcher built from split rules to find longest matched substrings, and add corresponding edges to the graph. wordcut-engine removes edges that break clusters. The example of split rules are shown below:
[\r\t\n ]+
[A-Za-z]+
[0-9]+
[๐-๙]+
[\(\)"'`\[\]{}\\/]
If there is no edge for each of character indice yet, a default maker create a edge that connected the known rightmost boundary.
Dependencies
~3.5–5MB
~104K SLoC