Data struct with codepoint builder for radicals (#7805)

Issue: #6941 

Modified from #7722 and #7646 

## Changelog
- `icu_segmenter`: Add unstable Unihan radical provider data and baked
support
* New types: `icu_segmenter::provider::UnihanIrgData<'data>`,
`icu_segmenter::provider::SegmenterUnihanRadicalV1`
* New associated const:
`icu_segmenter::provider::Baked::SINGLETON_SEGMENTER_UNIHAN_RADICAL_V1`
* The `experimental_segmenter` example now uses `radaboost` for the
Chinese radical model and adds `thadaboost` for Thai
- `icu_provider_source`: Add Unihan radical trie generation for
`icu_segmenter::provider::SegmenterUnihanRadicalV1`
  * `SourceDataProvider` can now load this marker from Unihan IRG data
- `icu_provider_registry`: Export
`icu_segmenter::provider::SegmenterUnihanRadicalV1`

Author: opnuub

components/icu/Cargo.toml

@@ -128,6 +128,7 @@ unstable = [
     "icu_calendar/unstable",
     "icu_datetime/unstable",
     "icu_plurals/unstable",
+    "icu_segmenter/unstable",
     "icu_time/unstable",
     "dep:icu_experimental",
     "dep:icu_pattern",

components/segmenter/Cargo.toml

@@ -56,6 +56,7 @@ default = ["compiled_data", "auto"]
 serde = ["dep:serde", "potential_utf/serde", "zerovec/serde", "icu_collections/serde", "icu_provider/serde"]
 datagen = ["serde", "dep:databake", "potential_utf/databake", "zerovec/databake", "icu_collections/databake", "icu_provider/export"]
 lstm = ["dep:core_maths"]
+unstable = []
 auto = ["lstm"] # Enables [try_]new_auto constructors
 compiled_data = ["dep:icu_segmenter_data", "dep:icu_locale", "icu_locale?/compiled_data", "icu_provider/baked"]
 

components/segmenter/examples/experimental_segmenter.rs

@@ -14,24 +14,48 @@ mod cnn;
 use adaboost::Predictor;
 use cnn::{CnnSegmenter, RawCnnData};
 use icu_segmenter::{options::WordBreakOptions, WordSegmenter, WordSegmenterBorrowed};
-use std::time::SystemTime;
+use std::time::Instant;
 
 const REPETITIONS: usize = 1000;
 
-fn main_adaboost(args: &[String]) {
+fn main_radaboost(args: &[String]) {
     let segmenter = Predictor::for_test();
     let s = &args[0];
-    let start_time = SystemTime::now();
+    let start_time = Instant::now();
     for _ in 0..REPETITIONS {
         segmenter.predict(s);
     }
-    let elapsed = start_time.elapsed().unwrap();
+    let elapsed = start_time.elapsed();
     println!("Output:");
     let mut prev = 0;
     for breakpoint in segmenter.predict_breakpoints(s) {
         print!("{}|", &s[prev..breakpoint]);
         prev = breakpoint;
     }
+    if prev < s.len() {
+        print!("{}", &s[prev..]);
+    }
+    println!();
+    println!("{} repetitions done in: {:?}", REPETITIONS, elapsed);
+}
+
+fn main_thadaboost(args: &[String]) {
+    let segmenter = Predictor::for_test_thai();
+    let s = &args[0];
+    let start_time = Instant::now();
+    for _ in 0..REPETITIONS {
+        segmenter.predict_thai(s);
+    }
+    let elapsed = start_time.elapsed();
+    println!("Output:");
+    let mut prev = 0;
+    for breakpoint in segmenter.predict_thai_breakpoints(s) {
+        print!("{}|", &s[prev..breakpoint]);
+        prev = breakpoint;
+    }
+    if prev < s.len() {
+        print!("{}", &s[prev..]);
+    }
     println!();
     println!("{} repetitions done in: {:?}", REPETITIONS, elapsed);
 }
@@ -55,11 +79,11 @@ fn main_cnn(args: &[String]) {
         .unwrap();
     let segmenter = CnnSegmenter::new(&cnndata);
     let s = &args[0];
-    let start_time = SystemTime::now();
+    let start_time = Instant::now();
     for _ in 0..REPETITIONS {
         segmenter.segment_str(s);
     }
-    let elapsed = start_time.elapsed().unwrap();
+    let elapsed = start_time.elapsed();
     println!("Output:");
     let mut prev = 0;
     for breakpoint in segmenter.segment_str(s).to_breakpoints() {
@@ -82,11 +106,11 @@ fn main_lstm(mut args: &[String]) {
 }
 
 fn run_word_segmenter(segmenter: WordSegmenterBorrowed, s: &str) {
-    let start_time = SystemTime::now();
+    let start_time = Instant::now();
     for _ in 0..REPETITIONS {
         segmenter.segment_str(s).count(); // consume the iterator
     }
-    let elapsed = start_time.elapsed().unwrap();
+    let elapsed = start_time.elapsed();
     println!("Output:");
     let mut prev = 0;
     for breakpoint in segmenter.segment_str(s) {
@@ -104,7 +128,8 @@ fn main() {
         return;
     }
     match args[1].as_str() {
-        "adaboost" => main_adaboost(&args[2..]),
+        "radaboost" => main_radaboost(&args[2..]),
+        "thadaboost" => main_thadaboost(&args[2..]),
         "dict" | "dictionary" => main_dict(&args[2..]),
         "cnn" => main_cnn(&args[2..]),
         "lstm" => main_lstm(&args[2..]),

components/segmenter/src/provider/mod.rs

@@ -17,6 +17,8 @@
 
 mod lstm;
 pub use lstm::*;
+#[cfg(feature = "unstable")]
+pub mod radical;
 
 use crate::options::WordType;
 use icu_collections::codepointtrie::CodePointTrie;
@@ -51,6 +53,8 @@ const _: () = {
     impl_segmenter_break_line_v1!(Baked);
     #[cfg(feature = "lstm")]
     impl_segmenter_lstm_auto_v1!(Baked);
+    #[cfg(feature = "unstable")]
+    impl_segmenter_unihan_radical_v1!(Baked);
     impl_segmenter_break_word_v1!(Baked);
     impl_segmenter_break_word_override_v1!(Baked);
     impl_segmenter_break_sentence_override_v1!(Baked);
@@ -135,6 +139,8 @@ pub const MARKERS: &[DataMarkerInfo] = &[
     SegmenterDictionaryAutoV1::INFO,
     SegmenterDictionaryExtendedV1::INFO,
     SegmenterLstmAutoV1::INFO,
+    #[cfg(feature = "unstable")]
+    radical::SegmenterUnihanRadicalV1::INFO,
 ];
 
 /// Pre-processed Unicode data in the form of tables to be used for rule-based breaking.

components/segmenter/src/provider/radical.rs

@@ -0,0 +1,32 @@
+// This file is part of ICU4X. For terms of use, please see the file
+// called LICENSE at the top level of the ICU4X source tree
+// (online at: https://github.com/unicode-org/icu4x/blob/main/LICENSE ).
+
+//! Data provider struct definitions for radicals.
+
+use icu_collections::codepointtrie::CodePointTrie;
+use icu_provider::prelude::*;
+
+/// Data for Unihan radicals.
+#[derive(Debug, PartialEq, Clone, yoke::Yokeable, zerofrom::ZeroFrom)]
+#[cfg_attr(feature = "datagen", derive(serde::Serialize, databake::Bake))]
+#[cfg_attr(feature = "datagen", databake(path = icu_segmenter::provider::radical))]
+#[cfg_attr(feature = "serde", derive(serde::Deserialize))]
+pub struct UnihanRadicalsData<'data> {
+    /// Trie mapping code points to their IRG source radical ID (u8).
+    #[cfg_attr(feature = "serde", serde(borrow))]
+    pub trie: CodePointTrie<'data, u8>,
+}
+
+icu_provider::data_struct!(
+    UnihanRadicalsData<'_>,
+    #[cfg(feature = "datagen")]
+);
+
+icu_provider::data_marker!(
+    /// Marker for the singleton trie mapping code points to their Unihan IRG source radical IDs.
+    SegmenterUnihanRadicalV1,
+    "segmenter/unihan/radical/v1",
+    UnihanRadicalsData<'static>,
+    is_singleton = true
+);

components/segmenter/tests/adaboost/main.rs

@@ -4,59 +4,37 @@
 
 #![allow(dead_code)]
 
+use icu_segmenter::provider::{radical::UnihanRadicalsData, Baked};
 use std::collections::HashMap;
 
 static MODEL_FOR_TEST: &str = include_str!("model.json");
 static MODEL_FOR_TEST_THAI: &str = include_str!("model_thai.json");
 
-static CODEPOINTS: &[u16] = &[
-    20008, 20022, 20031, 20057, 20101, 20108, 20128, 20154, 20799, 20837, 20843, 20866, 20886,
-    20907, 20960, 20981, 20992, 21147, 21241, 21269, 21274, 21304, 21313, 21340, 21353, 21378,
-    21430, 21448, 21475, 22231, 22303, 22763, 22786, 22794, 22805, 22823, 22899, 23376, 23424,
-    23544, 23567, 23586, 23608, 23662, 23665, 24027, 24037, 24049, 24062, 24178, 24186, 24191,
-    24308, 24318, 24331, 24339, 24400, 24417, 24435, 24515, 25096, 25142, 25163, 25903, 25908,
-    25991, 26007, 26020, 26041, 26080, 26085, 26352, 26376, 26408, 27424, 27490, 27513, 27571,
-    27595, 27604, 27611, 27663, 27668, 27700, 28779, 29226, 29238, 29243, 29247, 29255, 29273,
-    29275, 29356, 29572, 29577, 29916, 29926, 29976, 29983, 29992, 30000, 30091, 30098, 30326,
-    30333, 30382, 30399, 30446, 30683, 30690, 30707, 31034, 31160, 31166, 31348, 31435, 31481,
-    31859, 31992, 32566, 32593, 32650, 32701, 32769, 32780, 32786, 32819, 32895, 32905, 33251,
-    33258, 33267, 33276, 33292, 33307, 33311, 33390, 33394, 33400, 34381, 34411, 34880, 34892,
-    34915, 35198, 35211, 35282, 35328, 35895, 35910, 35925, 35960, 35997, 36196, 36208, 36275,
-    36523, 36554, 36763, 36784, 36789, 37009, 37193, 37318, 37324, 37329, 38263, 38272, 38428,
-    38582, 38585, 38632, 38737, 38750, 38754, 38761, 38859, 38893, 38899, 38913, 39080, 39131,
-    39135, 39318, 39321, 39340, 39592, 39640, 39647, 39717, 39727, 39730, 39740, 39770, 40165,
-    40565, 40575, 40613, 40635, 40643, 40653, 40657, 40697, 40701, 40718, 40723, 40736, 40763,
-    40778, 40786, 40845, 40860, 40864,
-];
-
-pub(crate) fn get_radical(ch: char) -> u8 {
-    let id = ch as u32;
-
-    if !(19968..=40869).contains(&id) {
-        return 0;
-    }
-
-    let idx = CODEPOINTS.partition_point(|&b| (b as u32) <= id);
-    (idx as u8) + 1
+pub(crate) fn get_radical(radicals: &UnihanRadicalsData<'_>, ch: char) -> u8 {
+    radicals.trie.get(ch)
 }
 
-pub(crate) struct Predictor {
+pub(crate) struct Predictor<'a> {
     pub(crate) model: HashMap<String, HashMap<String, i16>>,
+    radicals: &'a UnihanRadicalsData<'a>,
 }
 
-impl Predictor {
-    pub(crate) fn from_json(json: &str) -> Self {
+impl<'a> Predictor<'a> {
+    pub(crate) fn from_json(json: &str, radicals: &'a UnihanRadicalsData<'a>) -> Self {
         let model: HashMap<String, HashMap<String, i16>> =
             serde_json::from_str(json).unwrap_or_default();
-        Self { model }
+        Self { model, radicals }
     }
 
     pub(crate) fn for_test() -> Self {
-        Self::from_json(MODEL_FOR_TEST)
+        Self::from_json(MODEL_FOR_TEST, Baked::SINGLETON_SEGMENTER_UNIHAN_RADICAL_V1)
     }
 
     pub(crate) fn for_test_thai() -> Self {
-        Self::from_json(MODEL_FOR_TEST_THAI)
+        Self::from_json(
+            MODEL_FOR_TEST_THAI,
+            Baked::SINGLETON_SEGMENTER_UNIHAN_RADICAL_V1,
+        )
     }
 
     pub(crate) fn predict(&self, sentence: &str) -> Vec<i16> {
@@ -73,15 +51,15 @@ impl Predictor {
 
             let mut score: i16 = 4;
 
-            let rad4 = get_radical(c);
+            let rad4 = get_radical(self.radicals, c);
             if rad4 != 0 {
                 if let Some(map) = self.model.get("RSRID") {
                     let key = format!("{}:{}", c_prev, rad4);
                     score += map.get(&key).copied().unwrap_or(0);
                 }
             }
 
-            let rad3 = get_radical(c_prev);
+            let rad3 = get_radical(self.radicals, c_prev);
             if rad3 != 0 {
                 if let Some(map) = self.model.get("LSRID") {
                     let key = format!("{}:{}", rad3, c);
@@ -247,6 +225,18 @@ impl Predictor {
         }
         breakpoints
     }
+
+    pub(crate) fn predict_thai_breakpoints(&self, sentence: &str) -> Vec<usize> {
+        let mut breakpoints = vec![0];
+        let mut offset = 0;
+        for (&score, ch) in self.predict_thai(sentence).iter().zip(sentence.chars()) {
+            offset += ch.len_utf8();
+            if score > 0 {
+                breakpoints.push(offset);
+            }
+        }
+        breakpoints
+    }
 }
 
 #[cfg(test)]

components/segmenter/tests/adaboost/python_test_output.txt

@@ -24,7 +24,7 @@
 2404
 -1604
 2824
-440
+320
 -525
 5881
 4892
@@ -46,4 +46,4 @@
 4605
 4324
 -2139
-5215
+5215

provider/data/segmenter/data/mod.rs

@@ -24,3 +24,4 @@ segmenter/lstm/auto/v1, und/Burmese_codepoints_exclusive_model4_heavy, 91365B, 9
 segmenter/lstm/auto/v1, und/Khmer_codepoints_exclusive_model4_heavy, 74665B, 74368B, cc6fe9f66fed196d
 segmenter/lstm/auto/v1, und/Lao_codepoints_exclusive_model4_heavy, 72160B, 71863B, 3f52a4025c7d618f
 segmenter/lstm/auto/v1, und/Thai_codepoints_exclusive_model4_heavy, 72327B, 72030B, 4486b38238d7c651
+segmenter/unihan/radical/v1, <singleton>, 13080B, 13020B, d875e702fbedc249