sudachi/analysis/lattice.rs
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
/*
* Copyright (c) 2021-2024 Works Applications Co., Ltd.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
use crate::analysis::inner::{Node, NodeIdx};
use crate::analysis::node::{LatticeNode, PathCost, RightId};
use crate::dic::connect::ConnectionMatrix;
use crate::dic::grammar::Grammar;
use crate::dic::lexicon_set::LexiconSet;
use crate::dic::subset::InfoSubset;
use crate::dic::word_id::WordId;
use crate::error::SudachiResult;
use crate::input_text::InputBuffer;
use crate::prelude::SudachiError;
use std::fmt::{Display, Formatter};
use std::io::Write;
/// Lattice Node for Viterbi Search.
/// Extremely small for better cache locality.
/// Current implementation has 25% efficiency loss because of padding :(
/// Maybe we should use array-of-structs layout instead, but I want to try to measure the
/// efficiency of that without the effects of the current rewrite.
struct VNode {
total_cost: i32,
right_id: u16,
}
impl RightId for VNode {
#[inline]
fn right_id(&self) -> u16 {
self.right_id
}
}
impl PathCost for VNode {
#[inline]
fn total_cost(&self) -> i32 {
self.total_cost
}
}
impl VNode {
#[inline]
fn new(right_id: u16, total_cost: i32) -> VNode {
VNode {
right_id,
total_cost,
}
}
}
/// Lattice which is constructed for performing the Viterbi search.
/// Contain several parallel arrays.
/// First level of parallel arrays is indexed by end word boundary.
/// Word boundaries are always aligned to codepoint boundaries, not to byte boundaries.
///
/// During the successive analysis, we do not drop inner vectors, so
/// the size of vectors never shrink.
/// You must use the size parameter to check the current size and never
/// access vectors after the end.
#[derive(Default)]
pub struct Lattice {
ends: Vec<Vec<VNode>>,
ends_full: Vec<Vec<Node>>,
indices: Vec<Vec<NodeIdx>>,
eos: Option<(NodeIdx, i32)>,
size: usize,
}
impl Lattice {
fn reset_vec<T>(data: &mut Vec<Vec<T>>, target: usize) {
for v in data.iter_mut() {
v.clear();
}
let cur_len = data.len();
if cur_len <= target {
data.reserve(target - cur_len);
for _ in cur_len..target {
data.push(Vec::with_capacity(16))
}
}
}
/// Prepare lattice for the next analysis of a sentence with the
/// specified length (in codepoints)
pub fn reset(&mut self, length: usize) {
Self::reset_vec(&mut self.ends, length + 1);
Self::reset_vec(&mut self.ends_full, length + 1);
Self::reset_vec(&mut self.indices, length + 1);
self.eos = None;
self.size = length + 1;
self.connect_bos();
}
fn connect_bos(&mut self) {
self.ends[0].push(VNode::new(0, 0));
}
/// Find EOS node -- finish the lattice construction
pub fn connect_eos(&mut self, conn: &ConnectionMatrix) -> SudachiResult<()> {
let len = self.size;
let eos_start = (len - 1) as u16;
let eos_end = (len - 1) as u16;
let node = Node::new(eos_start, eos_end, 0, 0, 0, WordId::EOS);
let (idx, cost) = self.connect_node(&node, conn);
if cost == i32::MAX {
Err(SudachiError::EosBosDisconnect)
} else {
self.eos = Some((idx, cost));
Ok(())
}
}
/// Insert a single node in the lattice, founding the path to the previous node
/// Assumption: lattice for all previous boundaries is already constructed
pub fn insert(&mut self, node: Node, conn: &ConnectionMatrix) -> i32 {
let (idx, cost) = self.connect_node(&node, conn);
let end_idx = node.end();
self.ends[end_idx].push(VNode::new(node.right_id(), cost));
self.indices[end_idx].push(idx);
self.ends_full[end_idx].push(node);
cost
}
/// Find the path with the minimal cost through the lattice to the attached node
/// Assumption: lattice for all previous boundaries is already constructed
#[inline]
pub fn connect_node(&self, r_node: &Node, conn: &ConnectionMatrix) -> (NodeIdx, i32) {
let begin = r_node.begin();
let node_cost = r_node.cost() as i32;
let mut min_cost = i32::MAX;
let mut prev_idx = NodeIdx::empty();
for (i, l_node) in self.ends[begin].iter().enumerate() {
if !l_node.is_connected_to_bos() {
continue;
}
let connect_cost = conn.cost(l_node.right_id(), r_node.left_id()) as i32;
let new_cost = l_node.total_cost() + connect_cost + node_cost;
if new_cost < min_cost {
min_cost = new_cost;
prev_idx = NodeIdx::new(begin as u16, i as u16);
}
}
(prev_idx, min_cost)
}
/// Checks if there exist at least one at the word end boundary
pub fn has_previous_node(&self, i: usize) -> bool {
self.ends.get(i).map(|d| !d.is_empty()).unwrap_or(false)
}
/// Lookup a node for the index
pub fn node(&self, id: NodeIdx) -> (&Node, i32) {
let node = &self.ends_full[id.end() as usize][id.index() as usize];
let cost = self.ends[id.end() as usize][id.index() as usize].total_cost;
(node, cost)
}
/// Fill the path with the minimum cost (indices only).
/// **Attention**: the path will be reversed (end to beginning) and will need to be traversed
/// in the reverse order.
pub fn fill_top_path(&self, result: &mut Vec<NodeIdx>) {
if self.eos.is_none() {
return;
}
// start with EOS
let (mut idx, _) = self.eos.unwrap();
result.push(idx);
loop {
let prev_idx = self.indices[idx.end() as usize][idx.index() as usize];
if prev_idx.end() != 0 {
// add if not BOS
result.push(prev_idx);
idx = prev_idx;
} else {
// finish if BOS
break;
}
}
}
}
impl Lattice {
pub fn dump<W: Write>(
&self,
input: &InputBuffer,
grammar: &Grammar,
lexicon: &LexiconSet,
out: &mut W,
) -> SudachiResult<()> {
enum PosData<'a> {
Bos,
Borrow(&'a [String]),
}
impl Display for PosData<'_> {
fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
match self {
PosData::Bos => write!(f, "BOS/EOS"),
PosData::Borrow(data) => {
for (i, s) in data.iter().enumerate() {
write!(f, "{}", s)?;
if i + 1 != data.len() {
write!(f, ", ")?;
}
}
Ok(())
}
}
}
}
let mut dump_idx = 0;
for boundary in (0..self.indices.len()).rev() {
for r_node in &self.ends_full[boundary] {
let (surface, pos) = if r_node.is_special_node() {
("(null)", PosData::Bos)
} else if r_node.is_oov() {
let pos_id = r_node.word_id().word() as usize;
(
input.curr_slice_c(r_node.begin()..r_node.end()),
PosData::Borrow(&grammar.pos_list[pos_id]),
)
} else {
let winfo =
lexicon.get_word_info_subset(r_node.word_id(), InfoSubset::POS_ID)?;
(
input.orig_slice_c(r_node.begin()..r_node.end()),
PosData::Borrow(&grammar.pos_list[winfo.pos_id() as usize]),
)
};
write!(
out,
"{}: {} {} {}{} {} {} {} {}:",
dump_idx,
r_node.begin(),
r_node.end(),
surface,
r_node.word_id(),
pos,
r_node.left_id(),
r_node.right_id(),
r_node.cost()
)?;
let conn = grammar.conn_matrix();
for l_node in &self.ends[r_node.begin()] {
let connect_cost = conn.cost(l_node.right_id(), r_node.left_id());
write!(out, " {}", connect_cost)?;
}
writeln!(out)?;
dump_idx += 1;
}
}
Ok(())
}
}