chaffjs 0.29.0 → 0.31.0

package/src/facts/unit-conflicts.ts

@@ -1,14 +1,16 @@
 import type { ScopedFact } from "./fact-scope.ts";
 import { sameMeasure, type Measured } from "./measures.ts";
+import { toleranceOf } from "./unit-tolerance.ts";
+import { precisionTolerance, writtenHalfStep } from "./written-precision.ts";
+import { approximateTolerance } from "./approximate-tolerance.ts";
 
 /**
  * 同じ名前の量を、違う単位で書いて、換算すると合わない（「距離：5 km」と「距離：3000 m」）。比べるのは同じ範囲の、同じ種類の量で、
- * 単位が違うものだけ。同じ単位どうしの食い違いは fact-conflict が見る。丸めた書き方（1 mile と 1.6 km）は、差が TOLERANCE 以内なら合う。
+ * 単位が違うものだけ。同じ単位どうしの食い違いは fact-conflict が見る。丸めた書き方（1 mile と 1.6 km）は、差が量の種類ごとの許す差（unit-tolerance）以内なら合う。
+ * 小数で書いた量は、書いた桁の半分までの差も合う（1.2 kg と 1,250 g）。目安の量（約1 kg）は整数でも書いた桁の半分まで合う。
  */
 export type UnitConflict = { readonly fact: ScopedFact; readonly measured: Measured; readonly other: Measured };
 
-const TOLERANCE = 0.02;
-
 const groupKey = (fact: ScopedFact): string => `${fact.scope}\u0000${fact.key}`;
 
 type Read = { readonly fact: ScopedFact; readonly measured: Measured };
@@ -17,22 +19,33 @@ type Read = { readonly fact: ScopedFact; readonly measured: Measured };
  * 一つの種類（長さなら長さ）の量のうち、前にある単位の違う量のどれとも換算して合わないもの。どれか一つと合えば、食い違いは同じ単位
  * どうし（3 km と 5 km）の側にあり、fact-conflict が見る。
  */
-const conflictsInDimension = (reads: readonly Read[]): UnitConflict[] =>
+const conflictsInDimension = (reads: readonly Read[], halfStepOf: (measured: Measured) => number): UnitConflict[] =>
   reads.flatMap(({ fact, measured }, index) => {
     const others = reads.slice(0, index).filter((earlier) => earlier.measured.unit !== measured.unit);
     const [first] = others;
-    if (first === undefined || others.some((earlier) => sameMeasure(earlier.measured, measured, TOLERANCE))) return [];
+    const agrees = (earlier: Read): boolean => {
+      const written = precisionTolerance(toleranceOf(measured.dimension), [halfStepOf(earlier.measured), halfStepOf(measured)]);
+      const approximate = [earlier, { fact, measured }].filter((read) => read.fact.approximate === true).map((read) => read.measured);
+      return sameMeasure(earlier.measured, measured, approximateTolerance(written, approximate));
+    };
+    if (first === undefined || others.some(agrees)) return [];
     return [{ fact, measured, other: first.measured }];
   });
 
-const conflictsIn = (reads: readonly Read[]): UnitConflict[] => {
+const conflictsIn = (reads: readonly Read[], halfStepOf: (measured: Measured) => number): UnitConflict[] => {
   if (reads.some(({ fact }) => fact.record)) return [];
   const dimensions = [...new Set(reads.map(({ measured }) => measured.dimension))];
-  return dimensions.flatMap((dimension) => conflictsInDimension(reads.filter(({ measured }) => measured.dimension === dimension)));
+  return dimensions.flatMap((dimension) =>
+    conflictsInDimension(
+      reads.filter(({ measured }) => measured.dimension === dimension),
+      halfStepOf,
+    ),
+  );
 };
 
 /** 範囲と名前ごとに、最初の量と単位の違う量を換算して比べる。 */
-export const unitConflicts = (facts: readonly ScopedFact[], measuredOf: (fact: ScopedFact) => Measured | undefined): UnitConflict[] => {
+export const unitConflicts = (facts: readonly ScopedFact[], measuredOf: (fact: ScopedFact) => Measured | undefined, source: string): UnitConflict[] => {
+  const halfStepOf = (measured: Measured): number => writtenHalfStep(source.slice(measured.start, measured.end), measured.factors);
   const groups = new Map<string, Read[]>();
   facts.forEach((fact) => {
     const measured = measuredOf(fact);
@@ -41,5 +54,5 @@ export const unitConflicts = (facts: readonly ScopedFact[], measuredOf: (fact: S
     if (group === undefined) groups.set(groupKey(fact), [{ fact, measured }]);
     else group.push({ fact, measured });
   });
-  return [...groups.values()].flatMap(conflictsIn);
+  return [...groups.values()].flatMap((reads) => conflictsIn(reads, halfStepOf));
 };

package/src/facts/unit-equivalents.ts

@@ -0,0 +1,70 @@
+import { sameMeasure, sameUnit, type Measured, type Tolerance } from "./measures.ts";
+import { toleranceOf } from "./unit-tolerance.ts";
+
+/**
+ * 一つの量を二つの単位で並べて書いた所（180℃（350°F）、300 kPa (about 44 psi)）。括弧の中の量は、外の量を換算して丸めたもの。
+ * 括弧は量のすぐ後ろ（空白一つまで）で、中は前置き（約、about）を一つまで挟んですぐ量。量の後ろに字や数が続く括弧は、換算ではない。
+ * どちらが元の量かは分からないので、二つの数をそれぞれの桁で丸めたものとして比べる。
+ */
+export type UnitPair = { readonly measured: Measured; readonly other: Measured };
+
+const OPENERS = ["（", "("];
+const LETTER_OR_DIGIT = /[\p{L}\p{N}]/u;
+const LATIN_LETTER = /\p{Script=Latin}/u;
+const DECIMALS = /[.．]([0-9０-９]+)/u;
+const DIGITS = /[0-9０-９]+/u;
+const HALF = 0.5;
+const DECIMAL_BASE = 10;
+
+const skipSpace = (source: string, offset: number): number => (source.charAt(offset) === " " ? offset + 1 : offset);
+
+/** 前置きの後ろ。英字で終わる前置き（about）は、後ろに英字が続けば前置きではない（aboutness）。 */
+const afterHedge = (source: string, offset: number, hedges: readonly string[]): number => {
+  const rest = source.slice(offset).toLowerCase();
+  const hedge = hedges.find(
+    (word) => rest.startsWith(word.toLowerCase()) && !(LATIN_LETTER.test(word.slice(-1)) && LATIN_LETTER.test(rest.charAt(word.length))),
+  );
+  return hedge === undefined ? offset : skipSpace(source, offset + hedge.length);
+};
+
+/** 量の後ろの括弧の中で、換算した量が始まる所。括弧が無ければ undefined。 */
+const insideBracket = (source: string, end: number, hedges: readonly string[]): number | undefined => {
+  const open = skipSpace(source, end);
+  return OPENERS.includes(source.charAt(open)) ? afterHedge(source, open + 1, hedges) : undefined;
+};
+
+/**
+ * 書いた数を丸めた桁の半分を、基準の単位で（44 psi なら 0.5 psi、1.6 km なら 0.05 km）。0 で終わる整数は、最後の 0 の一つ上の桁で
+ * 丸めたとみなす（100 miles は 10 の位で ±5）。0 一つ（60、180）は 1 の位のまま。
+ */
+const trailingZeros = (digits: string): number => {
+  const reversed = [...digits.normalize("NFKC")].reverse();
+  const nonZero = reversed.findIndex((digit) => digit !== "0");
+  return nonZero === -1 ? reversed.length : nonZero;
+};
+
+const halfRoundingStep = (source: string, measured: Measured): number => {
+  const written = source.slice(measured.start, measured.end).replace(/,/gu, "");
+  const decimals = DECIMALS.exec(written)?.[1]?.length;
+  const zeros = trailingZeros(DIGITS.exec(written)?.[0] ?? "");
+  const exponent = decimals === undefined ? Math.max(0, zeros - 1) : -decimals;
+  return HALF * DECIMAL_BASE ** exponent * Math.min(...measured.factors);
+};
+
+/** 二つの数が、同じ値をそれぞれの桁で丸めたものでありうるなら合う。温度は、オーブンの目盛りの差（unit-tolerance）までも合う。 */
+const pairTolerance = (source: string, first: Measured, second: Measured): Tolerance => {
+  const tolerance = toleranceOf(second.dimension);
+  return { relative: tolerance.relative, absolute: Math.max(tolerance.absolute, halfRoundingStep(source, first) + halfRoundingStep(source, second)) };
+};
+
+/** 二つの単位で並べた量のうち、換算すると合わないもの。 */
+export const unitPairs = (source: string, measured: readonly Measured[], hedges: readonly string[]): UnitPair[] => {
+  const byStart = new Map(measured.map((value) => [value.start, value]));
+  return measured.flatMap((other): UnitPair[] => {
+    const start = insideBracket(source, other.end, hedges);
+    const value = start === undefined ? undefined : byStart.get(start);
+    if (value === undefined || value.dimension !== other.dimension || sameUnit(value, other)) return [];
+    if (LETTER_OR_DIGIT.test(source.charAt(value.end))) return [];
+    return sameMeasure(other, value, pairTolerance(source, other, value)) ? [] : [{ measured: value, other }];
+  });
+};

package/src/facts/unit-tolerance.ts

@@ -0,0 +1,17 @@
+import type { Tolerance } from "./measures.ts";
+
+/** 丸めた書き方（1 mile と 1.6 km、300 kPa と 44 psi）を合うとみなす差。大きいほうの値の 2%。 */
+const ROUNDING_SHARE = 0.02;
+
+/**
+ * 温度は割合では比べられない（0 ℃ の 2% は 0）。オーブンの換算表は 25 °F 刻みの目盛りに丸める（170 ℃ と 325 °F は 13 °F 違う）ので、
+ * 一目盛りより小さい 15 °F までの差は合うとみなす。
+ */
+const OVEN_ROUNDING_FAHRENHEIT = 15;
+const CELSIUS_PER_FAHRENHEIT_DEGREE = 5 / 9;
+
+const DEFAULT_TOLERANCE: Tolerance = { relative: ROUNDING_SHARE, absolute: 0 };
+const TEMPERATURE_TOLERANCE: Tolerance = { relative: 0, absolute: OVEN_ROUNDING_FAHRENHEIT * CELSIUS_PER_FAHRENHEIT_DEGREE };
+
+/** 量の種類（語彙表の id）ごとの、換算して合うとみなす差。 */
+export const toleranceOf = (dimension: string): Tolerance => (dimension === "unit-temperature" ? TEMPERATURE_TOLERANCE : DEFAULT_TOLERANCE);

package/src/facts/unit-word.ts

@@ -0,0 +1,35 @@
+const LATIN_WORD = /^\p{Script=Latin}+$/u;
+const WORD_CHAR = /[\p{L}\p{N}]/u;
+/** 単位のすぐ後ろに、数（空白一つまで）か、ハイフンでつながる語が続く。 */
+const CONTINUED = /^(?:[ \t]?[0-9０-９]|-[\p{L}\p{N}])/u;
+
+const isLatinWord = (word: string): boolean => LATIN_WORD.test(word);
+
+/** 英字の語が、語として行にあるか（"x" は "box" の中では数えない）。 */
+const hasWord = (line: string, word: string): boolean => {
+  let from = line.indexOf(word);
+  while (from !== -1) {
+    const end = from + word.length;
+    if (!WORD_CHAR.test(line.charAt(from - 1)) && !WORD_CHAR.test(line.charAt(end))) return true;
+    from = line.indexOf(word, from + 1);
+  }
+  return false;
+};
+
+/** 単位に要る文脈の語が、行にどれかあるか。英字の語は語として探し、ほかの字（オーブン、×）は字の並びとして探す。大文字と小文字は区別しない。 */
+export const lineHasContext = (line: string, words: readonly string[]): boolean => {
+  const lowered = line.toLowerCase();
+  return words.some((word) => {
+    const target = word.toLowerCase();
+    return isLatinWord(target) ? hasWord(lowered, target) : lowered.includes(target);
+  });
+};
+
+/**
+ * 文脈の要る英字の単位（"in"）は、その言語のふつうの語でもある。単位として読むのは、語がそこで終わり（"5 interns" の in ではない）、
+ * 後ろに数もハイフンでつながる語も続かない（"1 in 3"、"2-in-1"）ときだけ。文脈の要らない単位と、英字でない単位（度）は、この形を見ない。
+ */
+export const standsAsUnitWord = (pattern: string, needsContext: boolean, after: string): boolean => {
+  if (!needsContext || !isLatinWord(pattern)) return true;
+  return !WORD_CHAR.test(after.charAt(0)) && !CONTINUED.test(after);
+};

package/src/facts/written-precision.ts

@@ -0,0 +1,23 @@
+import type { Tolerance } from "./measures.ts";
+
+const HALF = 0.5;
+const DECIMAL_BASE = 10;
+const DECIMALS = /[.．]([0-9０-９]+)/u;
+/** 端の値（1.2 kg と 1,250 g）が浮動小数の誤差で外れないための余り。 */
+const FLOAT_SLACK = 1 + 1e-9;
+
+/**
+ * 小数で書いた量が言える幅の半分を、基準の単位で（1.2 kg は 1.15〜1.25 kg なので 0.05 kg）。書き手は小数点の後ろの最後の桁で丸めて
+ * 書いている。整数（410 g、1 mile）は 0：末尾の 0 が丸めか（1,200 g）は書き方から分からず、仕様の整数はふつう書いたとおりの値なので、
+ * 量の種類ごとの差（unit-tolerance）だけで比べる。
+ */
+export const writtenHalfStep = (written: string, factors: readonly number[]): number => {
+  const decimals = DECIMALS.exec(written)?.[1]?.length;
+  return decimals === undefined || factors.length === 0 ? 0 : HALF * DECIMAL_BASE ** -decimals * Math.max(...factors) * FLOAT_SLACK;
+};
+
+/** 二つの量を比べるときの、合うとみなす差。粗いほうの量の書いた桁の半分が、量の種類ごとの差より広ければそちら。 */
+export const precisionTolerance = (tolerance: Tolerance, halfSteps: readonly number[]): Tolerance => ({
+  relative: tolerance.relative,
+  absolute: Math.max(tolerance.absolute, ...halfSteps),
+});

package/src/kana-spelling.ts

@@ -22,6 +22,31 @@ export const kanjiSkeleton = (text: string): string => [...text.matchAll(KANJI)]
 export const dropsOkurigana = (spelling: string, spellings: readonly string[]): boolean =>
   spelling !== "" && kanjiSkeleton(spelling) === spelling && spellings.some((other) => other !== spelling && kanjiSkeleton(other) === spelling);
 
+const KANJI_OR_KANA_RUN = /[\p{Script=Han}々]|[^\p{Script=Han}々]+/gu;
+
+/**
+ * Whether a spelling writes in kana part of the word another spells with more kanji (かたくり粉 beside 片栗粉, とり扱い beside
+ * 取り扱い): every kanji it keeps stands where the other has it, each kana run stands for something there, and the other keeps
+ * more kanji. Only the letters are compared; the caller has already matched the readings.
+ */
+export const isKanaForKanji = (partial: string, full: string): boolean => {
+  const kept = kanjiSkeleton(partial);
+  if (kept === "" || kept === partial || [...kanjiSkeleton(full)].length <= [...kept].length) return false;
+  const pattern = [...partial.matchAll(KANJI_OR_KANA_RUN)].map(([part]) => (kanjiSkeleton(part) === part ? part : ".+")).join("");
+  return new RegExp(`^${pattern}$`, "u").test(full);
+};
+
+/**
+ * The kanji a spelling is keyed by among spellings of one reading: its own, or, when it writes part of the word in kana
+ * (かたくり粉), the kanji of the spelling with more kanji it stands for (片栗粉). When it could stand for two (two words of one
+ * reading), its own.
+ */
+export const kanjiKeyAmong = (spelling: string, spellings: readonly string[]): string => {
+  const fuller = new Set(spellings.filter((other) => isKanaForKanji(spelling, other)).map(kanjiSkeleton));
+  const [only] = fuller;
+  return fuller.size === 1 && only !== undefined ? only : kanjiSkeleton(spelling);
+};
+
 /** The length, in UTF-16 units, of the start both strings share, compared character by character (𠮟 is one character, two units). */
 const sharedPrefixLength = (left: string, right: string): number => {
   const [leftChars, rightChars] = [[...left], [...right]];

package/src/model-codes.ts

@@ -0,0 +1,104 @@
+// 型番の書き分け（KM-SP300 と KM-SP-300、KMSP300、km-sp300）。型番と読むのは、文書の題か型番の欄（語彙表 model-code-label）の後ろに
+// 書いた、英大文字と数字を合わせた語だけ。ハイフン・空白・大小・幅を除いて同じ字の並びになるものを同じ型番と見る。字や数の違う
+// 型番（KM-SP300 と KM-SP310）は別の型番なので比べない。
+
+export type ModelCode = { readonly offset: number; readonly surface: string };
+export type ModelCodeVariant = { readonly code: ModelCode; readonly usual: string };
+
+/** 型番と読む字の数（ハイフンと空白を除く）の下限。S3 や v2 は短すぎて型番か分からない。 */
+const MIN_KEY_LENGTH = 4;
+
+const ALNUM = "A-Za-z0-9Ａ-Ｚａ-ｚ０-９";
+const HYPHENS = "\\-‐‑－";
+/** 型番の字の間に書いてよい区切り。 */
+const SEPARATOR = `[${HYPHENS} ]?`;
+/** 型番の前に来れば、より長い語・パス・URL・ファイル名・色（#FF0000）の一部。 */
+const BEFORE_OTHER = new RegExp(`[${ALNUM}${HYPHENS}_./\\\\@=#]$`, "u");
+/** 型番の後ろに来れば、より長い語・パス・ファイル名（.pdf）・版（.2）の一部。 */
+const AFTER_OTHER = new RegExp(`^(?:[${ALNUM}${HYPHENS}_/\\\\]|\\.[${ALNUM}])`, "u");
+const CODE_SHAPE = new RegExp(`[${ALNUM}]+(?:[${HYPHENS}][${ALNUM}]+)*`, "gu");
+const URL = /(?:https?:\/\/|www\.)[^\s<>"'）)」]+/gu;
+
+/** ハイフン・空白・大小・幅を除いた字の並び。同じ並びなら同じ型番。 */
+export const modelCodeKey = (surface: string): string =>
+  surface
+    .normalize("NFKC")
+    .toLowerCase()
+    .replaceAll(/[\s\-‐‑]/gu, "");
+
+/** 英大文字と数字を両方含み、英小文字を含まない、短すぎない語。iOS17 や v1 は型番と読まない。 */
+export const isModelCodeShape = (surface: string): boolean => {
+  const key = modelCodeKey(surface);
+  const ascii = surface.normalize("NFKC");
+  return key.length >= MIN_KEY_LENGTH && /[A-Z]/u.test(ascii) && /\d/u.test(ascii) && !/[a-z]/u.test(ascii);
+};
+
+const standsAlone = (text: string, start: number, end: number): boolean =>
+  !BEFORE_OTHER.test(text.slice(Math.max(0, start - 1), start)) && !AFTER_OTHER.test(text.slice(end, end + 2));
+
+/** 範囲の中の、型番の形をした語。 */
+export const codesInTitle = (text: string, start: number, end: number): ModelCode[] =>
+  [...text.slice(start, end).matchAll(CODE_SHAPE)]
+    .map((match) => ({ offset: start + match.index, surface: match[0] }))
+    .filter((code) => isModelCodeShape(code.surface) && standsAlone(text, code.offset, code.offset + code.surface.length));
+
+const escaped = (pattern: string): string => pattern.replaceAll(/[.*+?^${}()|[\]\\]/gu, "\\$&");
+
+/** 欄の語の直後（「：」や空白を挟んでよい）に書いた、型番の形をした語。英字の欄の語は大小を問わず、語の途中からは読まない。 */
+export const codesAfterLabels = (text: string, labels: readonly string[]): ModelCode[] => {
+  if (labels.length === 0) return [];
+  const alternatives = labels.toSorted((left, right) => right.length - left.length).map(escaped);
+  const pattern = new RegExp(
+    `(?<![A-Za-z])(?:${alternatives.join("|")})(?![A-Za-z])[ \\t]*[:：]?[ \\t\\u3000]*([${ALNUM}]+(?:[${HYPHENS}][${ALNUM}]+)*)`,
+    "giu",
+  );
+  return [...text.matchAll(pattern)]
+    .map((match) => ({ offset: match.index + match[0].length - (match[1] ?? "").length, surface: match[1] ?? "" }))
+    .filter((code) => isModelCodeShape(code.surface) && standsAlone(text, code.offset, code.offset + code.surface.length));
+};
+
+/** 一つの字の、大小と全角の形。 */
+const charClass = (char: string): string => {
+  const forms = new Set([char, char.toUpperCase(), char.toLowerCase()].flatMap((form) => [form, String.fromCodePoint((form.codePointAt(0) ?? 0) + 0xfee0)]));
+  return `[${[...forms].join("")}]`;
+};
+
+const urlSpans = (text: string): { start: number; end: number }[] =>
+  [...text.matchAll(URL)].map((match) => ({ start: match.index, end: match.index + match[0].length }));
+
+/** 字の並びが key と同じになる語のすべて。字の間にはハイフンか空白を一つまで挟んでよい。URL の中は読まない。 */
+export const codesWithKey = (text: string, key: string): ModelCode[] => {
+  const pattern = new RegExp([...key].map(charClass).join(SEPARATOR), "gu");
+  const urls = urlSpans(text);
+  return [...text.matchAll(pattern)]
+    .map((match) => ({ offset: match.index, surface: match[0] }))
+    .filter((code) => standsAlone(text, code.offset, code.offset + code.surface.length))
+    .filter((code) => !urls.some((url) => url.start <= code.offset && code.offset < url.end));
+};
+
+const countsOf = (codes: readonly ModelCode[]): Map<string, number> =>
+  codes.reduce((counts, code) => counts.set(code.surface, (counts.get(code.surface) ?? 0) + 1), new Map<string, number>());
+
+/** 一番多く書いた形。数が同じなら、題か欄に先に書いた形。 */
+const usualOf = (codes: readonly ModelCode[], anchor: string): string => {
+  const counts = countsOf(codes);
+  const most = Math.max(...counts.values());
+  return (counts.get(anchor) ?? 0) === most ? anchor : ([...counts].find(([, count]) => count === most)?.[0] ?? anchor);
+};
+
+/** 一つの型番の、少ないほうの形。型番が二度以上出てこなければ、比べるものが無い。 */
+const variantsOfKey = (text: string, anchor: ModelCode): ModelCodeVariant[] => {
+  const codes = codesWithKey(text, modelCodeKey(anchor.surface));
+  if (codes.length < 2) return [];
+  const usual = usualOf(codes, anchor.surface);
+  return codes.filter((code) => code.surface !== usual).map((code) => ({ code, usual }));
+};
+
+/**
+ * 文書の型番（題か欄に書いたもの）を、ほかの所で別の形に書いた所。anchors は文書の順。同じ字の並びの型番は一度だけ見る。
+ */
+export const modelCodeVariants = (text: string, anchors: readonly ModelCode[]): ModelCodeVariant[] => {
+  const keys = anchors.map((anchor) => modelCodeKey(anchor.surface));
+  const firstPerKey = anchors.filter((_, index) => keys.indexOf(keys[index] ?? "") === index);
+  return firstPerKey.flatMap((anchor) => variantsOfKey(text, anchor)).toSorted((left, right) => left.code.offset - right.code.offset);
+};

package/src/name-spelling-chars.ts

@@ -0,0 +1,34 @@
+import type { NameMention } from "./name-variants.ts";
+
+// 一語の名前の中で、同じ音を二通りに書く字（桜ヶ丘・桜ケ丘・桜が丘、霞ヶ関・霞が関）。どの字が同じかは語彙表 name-spelling-char が
+// 組（group）ごとに言う。漢字に挟まれた字だけを寄せる。かなの語の中の字は寄せない。名前は解析器が一語と読んだもの
+// （霞が関）に限る。桜・が・丘 と読む書き方を本文にそのまま探すと、助詞の が（桜が丘を彩る）を名前と取り違える。
+
+/** 字から、その組の代表の字へ。 */
+export type SpellingChars = ReadonlyMap<string, string>;
+
+/**
+ * 名前の中の同じ音の字を比べる材料。chars はその字の組。alsoWritten は、この比べ方にだけ加える名前の現れ（表の升の名前）。升の名前をほかの見方に加えると、表の識別子（contentType）を本文の名前（Content-Type）と比べてしまう。
+ */
+export type SpellingInput = { readonly chars: SpellingChars; readonly alsoWritten: readonly NameMention[] };
+
+export const NO_SPELLING: SpellingInput = { chars: new Map(), alsoWritten: [] };
+
+const HAN = /^\p{Script=Han}$/u;
+
+const isBetweenHan = (chars: readonly string[], index: number): boolean => HAN.test(chars[index - 1] ?? "") && HAN.test(chars[index + 1] ?? "");
+
+/** 漢字に挟まれた、組のある字の位置。 */
+const foldableAt = (chars: readonly string[], spelling: SpellingChars): number[] =>
+  chars.flatMap((char, index) => (spelling.has(char) && isBetweenHan(chars, index) ? [index] : []));
+
+/** 漢字に挟まれた組の字を、組の代表に寄せた形。寄せる字が無ければ書いたまま。 */
+export const foldSpelling = (surface: string, spelling: SpellingChars): string => {
+  const chars = [...surface];
+  const at = new Set(foldableAt(chars, spelling));
+  return chars.map((char, index) => (at.has(index) ? (spelling.get(char) ?? char) : char)).join("");
+};
+
+/** 比べる形。漢字に挟まれた組の字を含む名前だけが、寄せた形を一つ持つ。含まない名前（Content-Type）はこの比べ方に入らない。 */
+export const foldedKeysOf = (surface: string, spelling: SpellingChars): string[] =>
+  foldableAt([...surface], spelling).length > 0 ? [foldSpelling(surface, spelling)] : [];

package/src/name-variants.ts

@@ -1,6 +1,8 @@
 import type { Token } from "./plugin.ts";
 import { escapeRegExp } from "./orthography.ts";
 import { nameCueAt, type NameCue, type NameCues } from "./name-cue.ts";
+import { isNearSurname } from "./surname-near.ts";
+import { foldedKeysOf, NO_SPELLING, type SpellingInput } from "./name-spelling-chars.ts";
 
 // 同じ名前（人・会社・製品）を、文書の中で少しだけ違う形に書いた所。どの形が正しいかは決めず、少ないほうを指す。
 // 三通りで同じ名前と見る。書き方の違いだけ（GitHub と Github、Mac OS と macOS）、読みが同じ（山田太郎 と 山田太朗）、
@@ -19,7 +21,11 @@ export type NameMention = {
   readonly cue?: NameCue;
 };
 
-export type NameVariant = { readonly mention: NameMention; readonly usual: string; readonly kind: "spelling" | "reading" | "near" | "character" };
+export type NameVariant = {
+  readonly mention: NameMention;
+  readonly usual: string;
+  readonly kind: "spelling" | "kana" | "reading" | "near" | "surname" | "character";
+};
 
 /** 字体の違う同じ字（斎・斉・齋）。字から、その組の代表の字へ。 */
 export type VariantChars = ReadonlyMap<string, string>;
@@ -99,6 +105,61 @@ export const mentionsIn = (tokens: readonly Token[], source: string, personSuffi
     return [{ surface, offset: first.span.start, reading, words: words.map((token) => token.surface), person }];
   });
 
+/** 隣の語を探す範囲（UTF-16 の単位）。 */
+const NEIGHBOUR_LOOKBACK = 40;
+/** 隣の語の外側の記号（(Rachel、Floyd.）。名前の側の記号（Rachel, Whitford の ,）は語を切るので残す。 */
+const MARK = /\p{P}/u;
+const withoutLeadingMarks = (word: string): string => {
+  const chars = [...word];
+  const start = chars.findIndex((char) => !MARK.test(char));
+  return start < 0 ? "" : chars.slice(start).join("");
+};
+const withoutTrailingMarks = (word: string): string => {
+  const chars = [...word];
+  return chars.slice(0, chars.findLastIndex((char) => !MARK.test(char)) + 1).join("");
+};
+
+const wordBeforeAt = (source: string, offset: number): string | undefined => {
+  if (source.charAt(offset - 1) !== " ") return undefined;
+  const from = Math.max(0, offset - 1 - NEIGHBOUR_LOOKBACK);
+  const parts = source.slice(from, offset - 1).split(/\s/u);
+  const word = withoutLeadingMarks(parts.at(-1) ?? "");
+  return word === "" || (from > 0 && parts.length === 1) ? undefined : word;
+};
+
+const wordAfterAt = (source: string, end: number): string | undefined => {
+  if (source.charAt(end) !== " ") return undefined;
+  const word = withoutTrailingMarks(source.slice(end + 1, end + 1 + NEIGHBOUR_LOOKBACK).split(/\s/u)[0] ?? "");
+  return word === "" ? undefined : word;
+};
+
+/**
+ * 解析器が名前の一語を名前と読まなかった所（to Rachel Whitford の Rachel を動詞と読む）。空白一つで隣の語と合わせた形を、ほかの所で
+ * 名前と読めていれば、その形の現れと見る。隣の語がほかの名前の現れに入るなら合わせない。
+ */
+export const withKnownNeighbours = (mentions: readonly NameMention[], source: string): NameMention[] => {
+  const known = new Set(mentions.filter((mention) => mention.words.length >= 2).map((mention) => mention.surface));
+  const taken = (start: number, end: number): boolean => mentions.some((mention) => mention.offset < end && start < mention.offset + mention.surface.length);
+  return mentions.map((mention) => {
+    const end = mention.offset + mention.surface.length;
+    const before = wordBeforeAt(source, mention.offset);
+    const after = wordAfterAt(source, end);
+    if (before !== undefined && known.has(`${before} ${mention.surface}`) && !taken(mention.offset - 1 - before.length, mention.offset - 1)) {
+      return {
+        ...mention,
+        surface: `${before} ${mention.surface}`,
+        offset: mention.offset - 1 - before.length,
+        reading: undefined,
+        words: [before, ...mention.words],
+      };
+    }
+    if (after !== undefined && known.has(`${mention.surface} ${after}`) && !taken(end + 1, end + 1 + after.length)) {
+      return { ...mention, surface: `${mention.surface} ${after}`, reading: undefined, words: [...mention.words, after] };
+    }
+    return mention;
+  });
+};
+
 /** 敬称の前の名前の字数（髙橋、髙橋太郎）。長い漢字の連なり（株式会社髙橋）は名前だけを切り出せない。 */
 const MAX_SUFFIXED_NAME = 4;
 const MIN_SUFFIXED_NAME = 2;
@@ -283,6 +344,20 @@ const isSlipOf = (tally: Tally, other: Tally): boolean => {
 
 const nearVariants = (tallies: readonly Tally[]): NameVariant[] => variantsIn(groupBy(tallies, maskedKeys), "near", isSlipOf);
 
+/**
+ * 名が同じで姓だけが二字ほど違う人の名前（Rachel Whitford と Rachel Whitfield）。一字違い（near）より遠いので、名と姓の二語以上
+ * で、違うのが最後の語のときに限る。相手が二度以上、こちらが一度だけ。名の違う二人（Sara Whitford と Rachel Whitfield）は比べない。
+ */
+const isSurnameSlipOf = (tally: Tally, other: Tally): boolean => {
+  if (tally.count !== 1 || other.count < 2) return false;
+  const [words, others] = [wordsOf(tally.surface), wordsOf(other.surface)];
+  const last = words.length - 1;
+  const sameBefore = words.slice(0, last).every((word, index) => word === others[index]);
+  return last >= 1 && words.length === others.length && sameBefore && isNearSurname(words[last] ?? "", others[last] ?? "");
+};
+
+const surnameVariants = (tallies: readonly Tally[]): NameVariant[] => variantsIn(groupBy(tallies, maskedKeys), "surname", isSurnameSlipOf);
+
 /** 字体の違う字（斎と斉）を組の代表の字に寄せた形。 */
 const characterKey = (tally: Tally, chars: VariantChars): string => [...tally.surface].map((letter) => chars.get(letter) ?? letter).join("");
 
@@ -331,28 +406,29 @@ const personReadingVariants = (tallies: readonly Tally[]): NameVariant[] =>
  * 同じ名前の、少ないほうの書き方。書き方ごとに最初の現れを一つ。同じ現れを二つの見方が言えば、先の見方だけ。chars は字体の
  * 違う同じ字の組（語彙表 name-variant-char）。
  */
-export const nameVariants = (mentions: readonly NameMention[], chars: VariantChars = new Map()): NameVariant[] => {
+export const nameVariants = (mentions: readonly NameMention[], chars: VariantChars = new Map(), spelling: SpellingInput = NO_SPELLING): NameVariant[] => {
   const tallies = talliesOf(mentions);
   const found = [
     ...variantsIn(
       groupBy(tallies, (tally) => [nameKey(tally.surface)]),
       "spelling",
     ),
+    ...variantsIn(
+      groupBy(talliesOf([...mentions, ...spelling.alsoWritten]), (tally) => foldedKeysOf(tally.surface, spelling.chars).map(nameKey)),
+      "kana",
+    ),
     ...variantsIn(
       groupBy(tallies, (tally) => (tally.first.reading === undefined ? [] : [tally.first.reading])),
       "reading",
       oneWordApart,
     ),
     ...nearVariants(tallies),
+    ...surnameVariants(tallies),
     ...characterVariants(tallies, chars),
     ...personReadingVariants(tallies),
   ];
   const reported = new Set<string>();
   return found
-    .filter((variant) => {
-      if (reported.has(variant.mention.surface)) return false;
-      reported.add(variant.mention.surface);
-      return true;
-    })
+    .filter((variant) => !reported.has(variant.mention.surface) && reported.add(variant.mention.surface).size > 0)
     .toSorted((left, right) => left.mention.offset - right.mention.offset);
 };