1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
use super::*;
use crate::ime::personal_rerank::allocations::count;
use crate::lattice::decode::tests::{lookup, row, syllables, table};
use crate::lattice::merge::block_buffer_tests::favouring_model;
use std::sync::Arc;
#[allow(clippy::too_many_arguments)]
fn baseline_merge_lattice_candidates(
candidates: &mut Vec<WordItem>,
syllables: &[String],
lookup: &mut LatticeLookup<'_>,
typed_pinyin: &str,
options: &LatticeOptions<'_>,
typo_source: Option<&mut TypoEdgeSource<'_>>,
rerankers: &mut [NeuralReranker],
rescoring_context: &str,
) -> Option<TypoSentence> {
if syllables.len() < 2 || !has_only_complete_pinyin_segments(syllables) {
return None;
}
let graph = build_graph(syllables, lookup, options);
let mut paths = decode_graph(&graph, options, None);
if paths.is_empty() {
return None;
}
let typo = match typo_source {
Some(source) if syllables.len() >= 3 => {
let edges = source(&paths[0]);
if edges.is_empty() {
None
} else {
decode_typo_on_graph(&graph, options, &paths[0], &edges)
}
}
_ => None,
};
if rerankers.is_empty() && options.emit > 0 {
paths.truncate(options.emit);
}
let dedup_size = candidates.len().saturating_add(paths.len());
let mut already =
(dedup_size > SMALL_SENTENCE_MERGE).then(|| HashSet::with_capacity(dedup_size));
if let Some(already) = already.as_mut() {
already.extend(candidates.iter().map(|item| item.word.as_str()));
}
let block = if rerankers.is_empty() {
if let Some(already) = already.as_ref() {
let mut block = Vec::new();
let capacity = paths.len();
for path in paths {
if already.contains(path.sentence.as_str()) {
continue;
}
if block.is_empty() {
block = Vec::with_capacity(capacity);
}
block.push(owned_sentence_row(
typed_pinyin,
path,
CandidateSource::Generated,
));
}
block
} else {
generated_block_linear(candidates, paths, typed_pinyin, options)
}
} else {
let mut keyboard = None;
for reranker in rerankers
.iter_mut()
.filter(|reranker| reranker.source == CandidateSource::NeuralKeyboard)
{
if let Some(order) = reranker.score_order(&paths, rescoring_context) {
keyboard = Some(order);
}
}
let path_count = paths.len();
let ranked = keyboard
.as_deref()
.map(|order| order.iter().copied().chain(order.len()..path_count));
if let Some(already) = already.as_ref() {
reranked_block(
paths,
0..path_count,
ranked,
options,
typed_pinyin,
|sentence| already.contains(sentence),
)
} else {
reranked_block(
paths,
0..path_count,
ranked,
options,
typed_pinyin,
|sentence| candidates.iter().any(|item| item.word == sentence),
)
}
};
if !block.is_empty() {
let at = whole_sentence_insert_position(candidates, syllables);
candidates.splice(at..at, block);
}
typo
}
#[test]
fn absent_sentence_sources_skip_only_the_unused_large_dedup_table() {
let keys = syllables("a'ba'ca");
assert!(has_only_complete_pinyin_segments(&keys));
let mut cases = 0;
let mut saved_cases = 0;
for n in [0, 1, 2, 6, 13, 65] {
let characters: Vec<_> = (0..n)
.map(|index| char::from_u32(0x4e00 + index as u32).unwrap())
.collect();
let mut rows = table(&[("ba", &[("吧", 1000)]), ("ca", &[("嚓", 1000)])]);
rows.insert(
"a".into(),
characters
.iter()
.enumerate()
.map(|(index, value)| row("a", &value.to_string(), 100_000 - index as i64))
.collect(),
);
let vocabulary: Vec<_> = characters
.iter()
.copied()
.chain(['吧', '嚓', '叭', '合', '成', '😀'])
.collect();
let model = Arc::new(favouring_model(
&vocabulary,
characters.last().map(std::slice::from_ref).unwrap_or(&[]),
));
for filler in [0, 62, 63, 64] {
for duplicates in [0, n] {
for include_lattice_best in [false, true] {
for show_next_on_duplicate in [false, true] {
for emit in [0, 1] {
let options = LatticeOptions {
nbest: n,
beam: n.max(32),
span_limit: n.max(32),
emit,
include_lattice_best,
show_next_on_duplicate,
..LatticeOptions::default()
};
let graph = build_graph(&keys, &mut lookup(&rows), &options);
let decoded = decode_graph(&graph, &options, None);
assert_eq!(decoded.len(), n);
let mut input: Vec<_> = decoded
.iter()
.take(duplicates)
.map(|path| {
WordItem::new(
"abaca",
&path.sentence,
123,
CandidateSource::Database,
"a'ba'ca",
)
})
.collect();
input.extend((0..filler).map(|index| {
WordItem::new(
"abaca",
format!("占位{index}"),
456,
CandidateSource::CloudSuggestion,
"",
)
}));
for mode in 0..4 {
let make_models = || match mode {
0 => vec![],
1 => vec![NeuralReranker::new(
CandidateSource::Generated,
model.clone(),
)],
2 => vec![NeuralReranker::new(
CandidateSource::NeuralKeyboard,
model.clone(),
)],
_ => vec![
NeuralReranker::new(
CandidateSource::NeuralKeyboard,
model.clone(),
),
NeuralReranker::new(
CandidateSource::Generated,
model.clone(),
),
NeuralReranker::new(
CandidateSource::NeuralKeyboard,
model.clone(),
),
],
};
for warm in [false, true] {
for typo_enabled in [false, true] {
let mut old_models = make_models();
let mut new_models = make_models();
let run =
|baseline: bool, models: &mut [NeuralReranker]| {
let mut candidates = input.clone();
let mut literal = None;
let mut source = |path: &SentencePath| {
literal = Some(path.clone());
vec![super::super::decode::TypoEdge {
start: 1,
end: 3,
key: "ba'ca".into(),
value: "叭嚓".into(),
weight: 900_000,
penalty: 0.5,
}]
};
let merge = if baseline {
baseline_merge_lattice_candidates
} else {
merge_lattice_candidates
};
let typo = merge(
&mut candidates,
&keys,
&mut lookup(&rows),
"abaca",
&options,
if typo_enabled {
Some(&mut source)
} else {
None
},
models,
if warm { "合成😀上下文" } else { "" },
);
(candidates, typo, literal)
};
if warm {
assert_eq!(
run(true, &mut old_models),
run(false, &mut new_models)
);
}
let (old, before) = count(|| run(true, &mut old_models));
let (new, after) = count(|| run(false, &mut new_models));
assert_eq!(new, old);
assert_eq!(new.0.capacity(), old.0.capacity());
let no_sources = mode > 0
&& !include_lattice_best
&& (mode == 1 || n < 2);
let saved = usize::from(
n > 0
&& no_sources
&& input.len() + n > SMALL_SENTENCE_MERGE,
);
let order_savings = if n < 2 {
0
} else {
usize::from(mode == 2) + 2 * usize::from(mode == 3)
};
// 有来源的单路径另外少建一张集合，无来源收益继续单独计量。
let paths = if mode == 0 && options.emit > 0 {
n.min(options.emit)
} else {
n
};
let single_path_savings = usize::from(
paths == 1
&& !no_sources
&& input.len() + paths > SMALL_SENTENCE_MERGE,
);
assert_eq!(after + saved + order_savings + single_path_savings, before, "路径={n}/占位={filler}/重复={duplicates}/字面={include_lattice_best}/继续={show_next_on_duplicate}/emit={emit}/模型={mode}/热={warm}/纠错={typo_enabled}");
if n > 0 && no_sources {
assert_eq!(new.0, input);
if typo_enabled {
assert!(new.1.is_some());
assert_eq!(new.2.as_ref().unwrap(), &decoded[0]);
}
}
for (old_model, new_model) in
old_models.iter_mut().zip(&mut new_models)
{
assert_eq!(
old_model.score_order(&decoded, "合成后续上下文"),
new_model.score_order(&decoded, "合成后续上下文")
);
}
saved_cases += saved;
cases += 1;
}
}
}
}
}
}
}
}
}
assert_eq!(cases, 6144);
assert_eq!(saved_cases, 544);
eprintln!("无来源去重表: {cases}组中{saved_cases}组精确少一次申请");
}