Cut input method startup from ~0.95 s to ~0.15 s
After login macOS starts the input method on first use, and keys typed while it loads its data were delayed or lost. Startup now: - parses lexicon.txt as raw bytes (no per-line String splitting or Double parsing) and builds the tone-ignoring index lazily on a background queue; - loads the precomputed English dictionary and bigram table (shared with the Windows build) instead of retraining the bigram model on every launch, and uses binary search instead of building a 230k-entry Set. Measured (release, warm cache): lexicon 0.65 -> 0.10 s, English 0.30 -> 0.04 s. The Windows (Python) core also builds its toneless index lazily: 0.56 -> 0.37 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
@@ -8,33 +8,63 @@ import Foundation
|
||||
/// - a letter-bigram model trained on those words, for words not in either list.
|
||||
public final class EnglishModel {
|
||||
private var common = Set<String>()
|
||||
private var dictionary = Set<String>()
|
||||
private var sortedWords: [String] = []
|
||||
private var sortedCommon: [String] = []
|
||||
/// Dictionary words (lowercase, > 2 letters), sorted: membership and prefix checks are
|
||||
/// binary searches, so no 230k-entry Set has to be built at startup.
|
||||
private var sortedDict: [String] = []
|
||||
private var bigram: [String: Double] = [:] // "ab" -> log2 P(b|a); "^a", "a$" for boundaries
|
||||
private var unseenBigram = -12.0
|
||||
|
||||
public init(commonWords: [String], dictionaryWords: [String] = []) {
|
||||
/// - Parameter bigram: precomputed table (english-bigram.json); trained from the words if nil.
|
||||
public init(commonWords: [String], dictionaryWords: [String] = [],
|
||||
bigram: (table: [String: Double], unseen: Double)? = nil) {
|
||||
common = Set(commonWords.map { $0.lowercased() }.filter { !$0.isEmpty })
|
||||
dictionary = Set(dictionaryWords.map { $0.lowercased() }.filter { $0.count > 2 })
|
||||
sortedCommon = common.sorted()
|
||||
sortedWords = common.union(dictionary).sorted()
|
||||
trainBigram()
|
||||
var dict = dictionaryWords.map { $0.lowercased() }.filter { $0.count > 2 }
|
||||
if !zip(dict, dict.dropFirst()).allSatisfy({ $0 < $1 }) {
|
||||
dict = Array(Set(dict)).sorted()
|
||||
}
|
||||
sortedDict = dict
|
||||
if let bigram {
|
||||
self.bigram = bigram.table
|
||||
unseenBigram = bigram.unseen
|
||||
} else {
|
||||
trainBigram()
|
||||
}
|
||||
}
|
||||
|
||||
/// Loads common words from a bundle resource file and the system dictionary if present.
|
||||
public convenience init(commonWordsURL: URL?, systemDictionary: URL? = URL(fileURLWithPath: "/usr/share/dict/words")) {
|
||||
/// Loads the bundled data files. Falls back to the system word list and to training
|
||||
/// the bigram table when the precomputed files are missing.
|
||||
public convenience init(commonWordsURL: URL?, dictionaryURL: URL? = nil, bigramURL: URL? = nil) {
|
||||
func lines(_ url: URL?) -> [String] {
|
||||
guard let url, let text = try? String(contentsOf: url, encoding: .utf8) else { return [] }
|
||||
return text.split(whereSeparator: \.isNewline)
|
||||
.map { $0.trimmingCharacters(in: .whitespaces) }
|
||||
.filter { !$0.hasPrefix("#") }
|
||||
.filter { !$0.isEmpty && !$0.hasPrefix("#") }
|
||||
}
|
||||
self.init(commonWords: lines(commonWordsURL), dictionaryWords: lines(systemDictionary))
|
||||
var bigram: (table: [String: Double], unseen: Double)?
|
||||
if let url = bigramURL, let data = try? Data(contentsOf: url),
|
||||
let obj = try? JSONSerialization.jsonObject(with: data) as? [String: Any],
|
||||
let table = obj["table"] as? [String: Double], let unseen = obj["unseen"] as? Double {
|
||||
bigram = (table, unseen)
|
||||
}
|
||||
let dictURL = dictionaryURL.flatMap { FileManager.default.fileExists(atPath: $0.path) ? $0 : nil }
|
||||
?? URL(fileURLWithPath: "/usr/share/dict/words")
|
||||
self.init(commonWords: lines(commonWordsURL), dictionaryWords: lines(dictURL), bigram: bigram)
|
||||
}
|
||||
|
||||
public func isCommon(_ w: String) -> Bool { common.contains(w.lowercased()) }
|
||||
public func isWord(_ w: String) -> Bool { let l = w.lowercased(); return common.contains(l) || dictionary.contains(l) }
|
||||
public func isWord(_ w: String) -> Bool { let l = w.lowercased(); return common.contains(l) || inDictionary(l) }
|
||||
|
||||
private func inDictionary(_ w: String) -> Bool {
|
||||
let i = lowerBound(w, in: sortedDict)
|
||||
return i < sortedDict.count && sortedDict[i] == w
|
||||
}
|
||||
|
||||
/// Some word strictly longer than `p` starts with it (common words or dictionary).
|
||||
private func isPrefixOfWord(_ p: String) -> Bool {
|
||||
hasPrefix(p, in: sortedCommon) || hasPrefix(p, in: sortedDict)
|
||||
}
|
||||
|
||||
/// - Parameter complete: true when the user ended the word (space / punctuation),
|
||||
/// false while still typing (prefixes of words get partial credit).
|
||||
@@ -43,10 +73,10 @@ public final class EnglishModel {
|
||||
let n = Double(w.count)
|
||||
guard !w.isEmpty else { return 0 }
|
||||
if common.contains(w) { return 3 + n }
|
||||
if dictionary.contains(w) { return 1 + 0.6 * n }
|
||||
if inDictionary(w) { return 1 + 0.6 * n }
|
||||
if !complete {
|
||||
if hasPrefix(w, in: sortedCommon) { return 0.4 * n }
|
||||
if hasPrefix(w, in: sortedWords) { return 0.2 * n }
|
||||
if isPrefixOfWord(w) { return 0.2 * n }
|
||||
}
|
||||
// Unknown word: judge by how English-like its letter sequence is.
|
||||
let plaus = plausibility(w)
|
||||
@@ -63,12 +93,12 @@ public final class EnglishModel {
|
||||
if common.contains(w) {
|
||||
f["lit.common"] = 1
|
||||
f["lit.common_len"] = n
|
||||
} else if dictionary.contains(w) {
|
||||
} else if inDictionary(w) {
|
||||
f["lit.dict"] = 1
|
||||
f["lit.dict_len"] = n
|
||||
} else if !complete && hasPrefix(w, in: sortedCommon) {
|
||||
f["lit.prefix_common_len"] = n
|
||||
} else if !complete && hasPrefix(w, in: sortedWords) {
|
||||
} else if !complete && isPrefixOfWord(w) {
|
||||
f["lit.prefix_dict_len"] = n
|
||||
} else {
|
||||
let plaus = plausibility(w)
|
||||
@@ -85,12 +115,17 @@ public final class EnglishModel {
|
||||
return total / Double(chars.count - 1)
|
||||
}
|
||||
|
||||
private func hasPrefix(_ p: String, in words: [String]) -> Bool {
|
||||
private func lowerBound(_ p: String, in words: [String]) -> Int {
|
||||
var lo = 0, hi = words.count
|
||||
while lo < hi {
|
||||
let mid = (lo + hi) / 2
|
||||
if words[mid] < p { lo = mid + 1 } else { hi = mid }
|
||||
}
|
||||
return lo
|
||||
}
|
||||
|
||||
private func hasPrefix(_ p: String, in words: [String]) -> Bool {
|
||||
var lo = lowerBound(p, in: words)
|
||||
// words[lo] is the first word >= p; skip an exact match, then test the next one.
|
||||
if lo < words.count && words[lo] == p { lo += 1 }
|
||||
return lo < words.count && words[lo].hasPrefix(p)
|
||||
@@ -99,7 +134,7 @@ public final class EnglishModel {
|
||||
private func trainBigram() {
|
||||
var pair: [String: Double] = [:]
|
||||
var first: [String: Double] = [:]
|
||||
for w in sortedWords where w.allSatisfy({ $0.isLetter && $0.isASCII }) {
|
||||
for w in Set(sortedCommon + sortedDict).sorted() where w.allSatisfy({ $0.isLetter && $0.isASCII }) {
|
||||
let weight = common.contains(w) ? 5.0 : 1.0
|
||||
let chars = ["^"] + w.map(String.init) + ["$"]
|
||||
for i in 1..<chars.count {
|
||||
|
||||
@@ -11,58 +11,159 @@ public struct LexiconEntry: Equatable {
|
||||
/// File format, one entry per line (`#` starts a comment):
|
||||
/// ㄋㄧˇ-ㄏㄠˇ 你好 -4.2
|
||||
/// The phrase/reading columns may be in either order; the score is optional.
|
||||
///
|
||||
/// Loading is on the input method's startup path, so the file is parsed as raw bytes,
|
||||
/// and the tone-ignoring index (only needed for untoned syllables) is built lazily —
|
||||
/// call `prepareTonelessIndex()` from a background queue to have it ready.
|
||||
public final class Lexicon {
|
||||
private var exact: [String: [LexiconEntry]] = [:]
|
||||
private var toneless: [String: [LexiconEntry]] = [:]
|
||||
private var tonelessIndex: [String: [LexiconEntry]]?
|
||||
private let lock = NSLock()
|
||||
public private(set) var maxPhraseLength = 1
|
||||
|
||||
public init() {}
|
||||
|
||||
public convenience init(contentsOf url: URL) throws {
|
||||
self.init()
|
||||
load(try String(contentsOf: url, encoding: .utf8))
|
||||
load(data: try Data(contentsOf: url))
|
||||
}
|
||||
|
||||
public func load(_ text: String) {
|
||||
for line in text.split(whereSeparator: \.isNewline) {
|
||||
let trimmed = line.trimmingCharacters(in: .whitespaces)
|
||||
if trimmed.isEmpty || trimmed.hasPrefix("#") { continue }
|
||||
let cols = trimmed.split(whereSeparator: { $0 == " " || $0 == "\t" }).map(String.init)
|
||||
guard cols.count >= 2 else { continue }
|
||||
let (reading, phrase) = Lexicon.isBopomofo(cols[0]) ? (cols[0], cols[1]) : (cols[1], cols[0])
|
||||
guard Lexicon.isBopomofo(reading) else { continue }
|
||||
let score = cols.count >= 3 ? Double(cols[2]) ?? -6 : -6
|
||||
add(reading: reading, phrase: phrase, score: score)
|
||||
}
|
||||
for key in exact.keys { exact[key]?.sort { $0.score > $1.score } }
|
||||
for key in toneless.keys { toneless[key]?.sort { $0.score > $1.score } }
|
||||
load(data: Data(text.utf8))
|
||||
}
|
||||
|
||||
public func add(reading rawReading: String, phrase: String, score: Double) {
|
||||
let entry = LexiconEntry(phrase: phrase, score: score)
|
||||
public func load(data: Data) {
|
||||
lock.lock()
|
||||
defer { lock.unlock() }
|
||||
exact.reserveCapacity(exact.count + data.count / 30)
|
||||
data.withUnsafeBytes { (raw: UnsafeRawBufferPointer) in
|
||||
let bytes = raw.bindMemory(to: UInt8.self)
|
||||
let n = bytes.count
|
||||
var i = 0
|
||||
while i < n {
|
||||
var end = i
|
||||
while end < n && bytes[end] != 0x0A { end += 1 }
|
||||
parseLine(bytes, i, end)
|
||||
i = end + 1
|
||||
}
|
||||
}
|
||||
for (key, entries) in exact where entries.count > 1 {
|
||||
exact[key] = entries.sorted { $0.score > $1.score }
|
||||
}
|
||||
tonelessIndex = nil // rebuilt on demand
|
||||
}
|
||||
|
||||
/// Splits one line into up to three space/tab separated fields.
|
||||
private func parseLine(_ b: UnsafeBufferPointer<UInt8>, _ start: Int, _ end: Int) {
|
||||
var fields: [Range<Int>] = []
|
||||
var i = start
|
||||
while i < end && fields.count < 3 {
|
||||
while i < end && (b[i] == 0x20 || b[i] == 0x09 || b[i] == 0x0D) { i += 1 }
|
||||
if i >= end { break }
|
||||
if fields.isEmpty && b[i] == 0x23 { return } // "#" comment
|
||||
let s = i
|
||||
while i < end && b[i] != 0x20 && b[i] != 0x09 && b[i] != 0x0D { i += 1 }
|
||||
fields.append(s..<i)
|
||||
}
|
||||
guard fields.count >= 2 else { return }
|
||||
func str(_ r: Range<Int>) -> String { String(decoding: UnsafeBufferPointer(rebasing: b[r]), as: UTF8.self) }
|
||||
let (r, p) = Lexicon.isBopomofo(b, fields[0]) ? (fields[0], fields[1]) : (fields[1], fields[0])
|
||||
guard Lexicon.isBopomofo(b, r) else { return }
|
||||
let score = fields.count >= 3 ? Lexicon.parseDouble(b, fields[2]) ?? -6 : -6
|
||||
// Neutral tone written as a prefix ("˙ㄉㄜ") needs normalising; detect it on bytes.
|
||||
var prefixNeutral = b[r.lowerBound] == 0xCB
|
||||
var k = r.lowerBound
|
||||
while !prefixNeutral && k + 2 < r.upperBound {
|
||||
if b[k] == 0x2D && b[k + 1] == 0xCB && b[k + 2] == 0x99 { prefixNeutral = true }
|
||||
k += 1
|
||||
}
|
||||
addEntry(reading: str(r), phrase: str(p), score: score, normalise: prefixNeutral)
|
||||
}
|
||||
|
||||
/// Parses "-3.9151"-style decimals without creating a String.
|
||||
private static func parseDouble(_ b: UnsafeBufferPointer<UInt8>, _ r: Range<Int>) -> Double? {
|
||||
var i = r.lowerBound
|
||||
var sign = 1.0
|
||||
if i < r.upperBound && (b[i] == 0x2D || b[i] == 0x2B) { sign = b[i] == 0x2D ? -1 : 1; i += 1 }
|
||||
var value = 0.0, scale = 1.0, digits = 0, fraction = false
|
||||
while i < r.upperBound {
|
||||
let c = b[i]
|
||||
if c >= 0x30 && c <= 0x39 {
|
||||
if fraction { scale /= 10; value += Double(c - 0x30) * scale } else { value = value * 10 + Double(c - 0x30) }
|
||||
digits += 1
|
||||
} else if c == 0x2E && !fraction {
|
||||
fraction = true
|
||||
} else {
|
||||
return Double(String(decoding: UnsafeBufferPointer(rebasing: b[r]), as: UTF8.self))
|
||||
}
|
||||
i += 1
|
||||
}
|
||||
return digits > 0 ? sign * value : nil
|
||||
}
|
||||
|
||||
/// Adds one entry (call `load` for bulk data; this does not re-sort).
|
||||
public func add(reading: String, phrase: String, score: Double) {
|
||||
lock.lock()
|
||||
defer { lock.unlock() }
|
||||
let key = addEntry(reading: reading, phrase: phrase, score: score, normalise: reading.contains("˙"))
|
||||
exact[key]?.sort { $0.score > $1.score }
|
||||
tonelessIndex = nil
|
||||
}
|
||||
|
||||
@discardableResult
|
||||
private func addEntry(reading rawReading: String, phrase: String, score: Double, normalise: Bool) -> String {
|
||||
var reading = rawReading
|
||||
// Some data sets write the neutral tone as a prefix ("˙ㄉㄜ"); we use a suffix.
|
||||
let syllables = rawReading.split(separator: "-").map { s -> String in
|
||||
s.hasPrefix("˙") ? String(s.dropFirst()) + "˙" : String(s)
|
||||
if normalise {
|
||||
reading = reading.split(separator: "-").map { s -> String in
|
||||
s.hasPrefix("˙") ? String(s.dropFirst()) + "˙" : String(s)
|
||||
}.joined(separator: "-")
|
||||
}
|
||||
let reading = syllables.joined(separator: "-")
|
||||
maxPhraseLength = max(maxPhraseLength, syllables.count)
|
||||
if !(exact[reading]?.contains { $0.phrase == phrase } ?? false) {
|
||||
exact[reading, default: []].append(entry)
|
||||
maxPhraseLength = max(maxPhraseLength, reading.utf8.reduce(1) { $0 + ($1 == 0x2D ? 1 : 0) })
|
||||
if exact[reading] == nil {
|
||||
exact[reading] = [LexiconEntry(phrase: phrase, score: score)]
|
||||
} else if !exact[reading]!.contains(where: { $0.phrase == phrase }) {
|
||||
exact[reading]!.append(LexiconEntry(phrase: phrase, score: score))
|
||||
}
|
||||
let key = Lexicon.stripTones(reading)
|
||||
if let i = toneless[key]?.firstIndex(where: { $0.phrase == phrase }) {
|
||||
toneless[key]![i].score = max(toneless[key]![i].score, score)
|
||||
} else {
|
||||
toneless[key, default: []].append(entry)
|
||||
return reading
|
||||
}
|
||||
|
||||
/// Builds the index that ignores tones (for untoned syllables). Safe to call from a
|
||||
/// background queue; lookups build it themselves if it is not ready yet.
|
||||
public func prepareTonelessIndex() {
|
||||
lock.lock()
|
||||
defer { lock.unlock() }
|
||||
_ = buildTonelessIfNeeded()
|
||||
}
|
||||
|
||||
private func buildTonelessIfNeeded() -> [String: [LexiconEntry]] {
|
||||
if let index = tonelessIndex { return index }
|
||||
var best: [String: [String: Double]] = [:]
|
||||
for (reading, entries) in exact {
|
||||
let key = Lexicon.stripTones(reading)
|
||||
for e in entries {
|
||||
if let s = best[key]?[e.phrase], s >= e.score { continue }
|
||||
best[key, default: [:]][e.phrase] = e.score
|
||||
}
|
||||
}
|
||||
var index: [String: [LexiconEntry]] = [:]
|
||||
index.reserveCapacity(best.count)
|
||||
for (key, phrases) in best {
|
||||
index[key] = phrases.map { LexiconEntry(phrase: $0.key, score: $0.value) }
|
||||
.sorted { $0.score != $1.score ? $0.score > $1.score : $0.phrase < $1.phrase }
|
||||
}
|
||||
tonelessIndex = index
|
||||
return index
|
||||
}
|
||||
|
||||
/// Entries for a sequence of readings. Untoned syllables match any tone.
|
||||
public func entries(for syllables: [Syllable]) -> [LexiconEntry] {
|
||||
lock.lock()
|
||||
defer { lock.unlock() }
|
||||
if syllables.allSatisfy({ $0.tone != nil }) {
|
||||
return exact[syllables.map(\.reading).joined(separator: "-")] ?? []
|
||||
}
|
||||
return toneless[syllables.map(\.toneless).joined(separator: "-")] ?? []
|
||||
return buildTonelessIfNeeded()[syllables.map(\.toneless).joined(separator: "-")] ?? []
|
||||
}
|
||||
|
||||
public func bestScore(for syllable: Syllable) -> Double? {
|
||||
@@ -81,4 +182,13 @@ public final class Lexicon {
|
||||
guard let first = s.unicodeScalars.first else { return false }
|
||||
return (0x3105...0x312F).contains(first.value) || first == "˙"
|
||||
}
|
||||
|
||||
/// Same test on raw UTF-8: Bopomofo U+3105–U+312F is E3 84 85…E3 84 AF; "˙" is CB 99.
|
||||
private static func isBopomofo(_ b: UnsafeBufferPointer<UInt8>, _ r: Range<Int>) -> Bool {
|
||||
guard r.count >= 2 else { return false }
|
||||
if b[r.lowerBound] == 0xCB && b[r.lowerBound + 1] == 0x99 { return true }
|
||||
guard r.count >= 3, b[r.lowerBound] == 0xE3, b[r.lowerBound + 1] == 0x84 else { return false }
|
||||
let c = b[r.lowerBound + 2]
|
||||
return c >= 0x85 && c <= 0xAF
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user