@whiteslove/parsing-lexicon 0.10.7 → 0.10.9

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@whiteslove/parsing-lexicon",
-  "version": "0.10.7",
+  "version": "0.10.9",
   "description": "Shared deterministic multilingual parsing lexicon for WhitesLove housing and hiring services",
   "repository": {
     "type": "git",

package/src/hiring-skills.js

@@ -368,8 +368,11 @@ export function canonicalSkillName(value) {
   return CANONICAL_BY_ALIAS.get(normalized) ?? extractSkillNames(normalized)[0]
 }
 
+const PAY_BONUS_KPI_RE = /(?:\+|plus|плюс)\s*(?:%\s*\/?\s*)?kpi(?![\p{L}\p{N}_])/giu
+
 export function extractSkillDetails(text) {
-  const normalized = normalizeSkillText(cleanHiringSourceText(text))
+  // "+ KPI" / "plus KPI" after a pay figure is a bonus scheme, not a skill.
+  const normalized = normalizeSkillText(cleanHiringSourceText(text).replace(PAY_BONUS_KPI_RE, ' '))
   const found = []
   const names = new Set()
   for (const { definition, patterns } of COMPILED_SKILLS) {

package/src/hiring-source-semantics.js

@@ -76,8 +76,8 @@ export function extractJobStructuredField(value, key, maxLength = 220) {
   return extractField(value, JOB_FIELD_TERMS[key], FIELD_EXTRA_ALIASES.job[key] || [], maxLength);
 }
 
-const INLINE_VACANCY_HEADING_RE = /✔️?\s*([A-ZА-ЯЁЎҚҒҲІЇЄ][A-ZА-ЯЁЎҚҒҲІЇЄ\s()\/-]{2,100}?)(?=\s*(?:[•⏳✔📍✈📞]|$))/gu;
-const SHARED_HIRING_CONTACT_RE = /(?:✈️?\s*Telegram\s*:|📞\s*Контакт\s*:)/iu;
+const INLINE_VACANCY_HEADING_RE = /(?:✔️?|▶️?)\s*([A-ZА-ЯЁЎҚҒҲІЇЄ][A-ZА-ЯЁЎҚҒҲІЇЄ\s()\/-]{2,100}?(?:\s*\([^)\n]{1,30}\))?)(?=\s*(?:[:•⏳✔▶📍✈📞💰]|$))/gu;
+const SHARED_HIRING_CONTACT_RE = /(?:✈️?\s*Telegram\s*:|📞\s*Контакт\s*:|👍\s*ГРАФИК\s+РАБОТЫ)/iu;
 
 /** Split a multi-vacancy source post into role-local blocks and shared context. */
 export function splitHiringVacancyEntries(value) {

package/src/location-merge.js

@@ -31,18 +31,29 @@ export const LOCATION_LIST_KEYS = Object.freeze([
   'searchClusters',
 ]);
 
+// One shared accessor compiles an entry's alias regex on first use. Giving
+// every entry its own getter closure cost ~100 bytes per entry, and the full
+// dictionary set holds over 130,000 entries. The compiled regex lives in a
+// WeakMap because entries are frozen.
+const COMPILED_RE = new WeakMap();
+const ENTRY_PROTOTYPE = Object.freeze({
+  get re() {
+    let re = COMPILED_RE.get(this);
+    if (!re) {
+      re = aliasesToRegex(this.aliases);
+      COMPILED_RE.set(this, re);
+    }
+    return re;
+  },
+});
+
+function frozenEntry(fields) {
+  return Object.freeze(Object.assign(Object.create(ENTRY_PROTOTYPE), fields));
+}
+
 export function locationEntry(name, ...aliases) {
   const all = [...new Set([name, ...aliases].flat().filter(Boolean))];
-  let re = null;
-  return Object.freeze({
-    canonical: name,
-    name,
-    aliases: Object.freeze(all),
-    get re() {
-      re ||= aliasesToRegex(all);
-      return re;
-    },
-  });
+  return frozenEntry({ canonical: name, name, aliases: Object.freeze(all) });
 }
 
 export function locationEntries(rows = []) {
@@ -61,6 +72,26 @@ function entryData(entry) {
   return data;
 }
 
+function uniqueTruthy(values) {
+  if (values.length > 32) return new Set(values).size === values.length && values.every(Boolean);
+  for (let i = 0; i < values.length; i += 1) {
+    if (!values[i]) return false;
+    for (let j = 0; j < i; j += 1) if (values[j] === values[i]) return false;
+  }
+  return true;
+}
+
+// Merging one entry with nothing returns an entry equal to itself when it is
+// already frozen, canonical, untyped and carries a unique alias list that
+// names it (every locationEntry()/mergeEntry() result does). Nearly every
+// group at import time is a single such entry, so reuse it instead of
+// rebuilding it. Map-data marking is by identity, so it is preserved as is.
+function isFinishedEntry(entry) {
+  if (!Object.isFrozen(entry) || !entry.canonical || entry.type !== undefined || entry.entityType !== undefined) return false;
+  const aliases = entry.aliases;
+  return Array.isArray(aliases) && Object.isFrozen(aliases) && aliases.includes(entry.name) && uniqueTruthy(aliases);
+}
+
 function mergeEntry(existing, incoming) {
   const aliases = [...new Set([
     ...(existing?.aliases || []),
@@ -73,19 +104,14 @@ function mergeEntry(existing, incoming) {
   // aliases. It remains fallback-only only when every merged source is map
   // data; otherwise the reviewed owner must retain its matching precedence.
   const mapData = existing ? isMapDataEntry(existing) && isMapDataEntry(incoming) : isMapDataEntry(incoming);
-  let re = null;
-  const result = {
+  const result = frozenEntry({
     ...base,
     canonical: base.canonical || base.name,
     type: base.type || base.entityType,
     aliases: Object.freeze(aliases),
-    get re() {
-      re ||= aliasesToRegex(aliases);
-      return re;
-    },
-  };
+  });
   if (mapData) MAP_DATA_ENTRIES.add(result);
-  return Object.freeze(result);
+  return result;
 }
 
 function parentKey(entry) {
@@ -111,6 +137,10 @@ export function mergeLocationEntries(...lists) {
   const result = [];
   for (const canonical of order) {
     const group = groups.get(canonical) || [];
+    if (group.length === 1) {
+      result.push(isFinishedEntry(group[0]) ? group[0] : mergeEntry(null, group[0]));
+      continue;
+    }
     const scopedParents = [...new Set(group.map(parentKey).filter(Boolean))];
 
     if (scopedParents.length <= 1) {

package/src/money-core.js

@@ -22,7 +22,10 @@ export const MONEY_SCALE_PATTERN = 'k|к|тыс\\.?|тысяч(?:а|и)?|тис\
 // present at all) would stop matching, since "до" doesn't satisfy the
 // boundary either.
 export const MONEY_RANGE_RE = new RegExp(`(${MONEY_NUMBER_PATTERN})\\s*(?:(${MONEY_SCALE_PATTERN})(?=$|[^\\p{L}\\p{N}_]))?\\s*(?:-|–|—|до|to|bis|dan\\s+gacha)\\s*(${MONEY_NUMBER_PATTERN})\\s*(?:(${MONEY_SCALE_PATTERN})(?=$|[^\\p{L}\\p{N}_]))?`, 'iu');
-export const MONEY_SINGLE_RE = new RegExp(`(${MONEY_NUMBER_PATTERN})\\s*(${MONEY_SCALE_PATTERN})?(?=$|[^\\p{L}\\p{N}_])`, 'giu');
+// Boards often glue the currency to the number ("10 000 000сум"); without this
+// the boundary check fails on the full number and backtracks to "10 000".
+const ATTACHED_CURRENCY = "сум|сўм|so['’ʻʼ`]?m|som|uzs|usd|eur|rub|руб|грн|kzt|kgs";
+export const MONEY_SINGLE_RE = new RegExp(`(${MONEY_NUMBER_PATTERN})\\s*(${MONEY_SCALE_PATTERN})?(?=$|[^\\p{L}\\p{N}_]|(?:${ATTACHED_CURRENCY})(?![\\p{L}\\p{N}_]))`, 'giu');
 
 export function parseNumericAmount(raw) {
   let value = String(raw || '').replace(/\u00a0/g, ' ').trim();

package/src/money.js

@@ -1,4 +1,4 @@
-import { findPhoneLikeSpans } from './contact.js';
+import { findPhoneLikeSpans, parsePhoneNumbers } from './contact.js';
 import { aliasesOf, findCanonical, normalizeUnicode } from './normalization.js';
 import {
   CURRENCY_SYMBOL_CANDIDATES,
@@ -56,7 +56,15 @@ function hasModifier(text, key) {
 function protectedPhoneSpans(text) {
   return findPhoneLikeSpans(text).filter((span) => {
     const prefix = text.slice(Math.max(0, span.start - 32), span.start);
-    return CONTACT_MARKER_RE.test(prefix);
+    if (CONTACT_MARKER_RE.test(prefix)) return true;
+    // A label-less international number ("998 953 506 090", typically scraped
+    // into a pay field) is a contact, not a 12-digit salary. Require a valid
+    // E.164 number of 11+ digits so ordinary grouped salaries stay money.
+    // A "998 95… - 998 95…" pair is merged into one span, so test each part.
+    return span.raw.split(/\s*[-–—]\s*/u).some((part) => {
+      const digits = part.replace(/\D/g, '');
+      return digits.length >= 11 && parsePhoneNumbers(`+${digits}`).length > 0;
+    });
   });
 }
 

package/src/normalization.js

@@ -10,8 +10,8 @@ const DASHES_RE = /[‐‑‒–—―]/g;
 // recent long inputs. Short strings (dictionary aliases, single words) are
 // cheaper to normalize than to cache and are left alone; the cap bounds
 // memory to a few recent texts per function.
-const MEMO_MIN_LENGTH = 200;
-const MEMO_MAX_ENTRIES = 32;
+const MEMO_MIN_LENGTH = 24;
+const MEMO_MAX_ENTRIES = 512;
 
 function memoizeLongText(fn) {
   const cache = new Map();
@@ -34,8 +34,15 @@ export const normalizeUnicode = memoizeLongText(function normalizeUnicode(value)
     .replace(DASHES_RE, '-');
 });
 
-/** Stable comparison form for parser dictionaries. */
-export const normalizeForMatch = memoizeLongText(function normalizeForMatch(value) {
+// Dictionary construction normalizes the same names and aliases hundreds of
+// thousands of times (about 5x more calls than distinct strings at import), so
+// short strings get a bounded cache of their own. Once it is full new strings
+// are simply computed, never evicted: the entries that fill it first are the
+// dictionary strings that every later match compares against.
+const SHORT_CACHE_MAX_ENTRIES = 131072;
+const shortMatchCache = new Map();
+
+function normalizeForMatchUncached(value) {
   return normalizeUnicode(value)
     .toLocaleLowerCase()
     .replace(/ё/g, 'е')
@@ -43,7 +50,19 @@ export const normalizeForMatch = memoizeLongText(function normalizeForMatch(valu
     .replace(/[^\p{L}\p{N}]+/gu, ' ')
     .replace(/\s+/g, ' ')
     .trim();
-});
+}
+
+const normalizeForMatchLong = memoizeLongText(normalizeForMatchUncached);
+
+/** Stable comparison form for parser dictionaries. */
+export function normalizeForMatch(value) {
+  if (typeof value !== 'string' || value.length >= MEMO_MIN_LENGTH) return normalizeForMatchLong(value);
+  const hit = shortMatchCache.get(value);
+  if (hit !== undefined) return hit;
+  const result = normalizeForMatchUncached(value);
+  if (shortMatchCache.size < SHORT_CACHE_MAX_ENTRIES) shortMatchCache.set(value, result);
+  return result;
+}
 
 /** Search-only apostrophe compaction for Uzbek/Karakalpak variants. */
 const normalizeCompactApostropheForMatch = memoizeLongText(function normalizeCompactApostropheForMatch(value) {
@@ -87,39 +106,50 @@ const KARAKALPAK_LATIN_PATTERN_EQUIVALENCE = Object.freeze({
   u: '[uú]', ú: '[uú]',
 });
 
+// Every map below is keyed by single BMP characters, so one regex replace is
+// identical to mapping the string character by character.
+function foldByMap(map) {
+  const re = new RegExp(`[${Object.keys(map).join('')}]`, 'g');
+  return (text) => text.replace(re, (char) => map[char]);
+}
+const foldCyrillicChars = foldByMap(CYRILLIC_SEARCH_MAP);
+const foldKazakhChars = foldByMap(KAZAKH_SEARCH_EQUIVALENCE);
+const foldKarakalpakChars = foldByMap(KARAKALPAK_LATIN_SEARCH_EQUIVALENCE);
+
 /** Search-oriented Cyrillic folding; canonical identity still comes from explicit aliases. */
 export const foldCyrillicForSearch = memoizeLongText(function foldCyrillicForSearch(value) {
-  return normalizeUnicode(value)
-    .toLocaleLowerCase()
-    .split('')
-    .map((char) => CYRILLIC_SEARCH_MAP[char] ?? char)
-    .join('');
+  return foldCyrillicChars(normalizeUnicode(value).toLocaleLowerCase());
 });
 
 const foldKazakhForSearch = memoizeLongText(function foldKazakhForSearch(value) {
-  return normalizeUnicode(value)
-    .toLocaleLowerCase()
-    .split('')
-    .map((char) => KAZAKH_SEARCH_EQUIVALENCE[char] ?? char)
-    .join('');
+  return foldKazakhChars(normalizeUnicode(value).toLocaleLowerCase());
 });
 
-function foldKarakalpakLatinForSearch(value) {
-  return normalizeUnicode(value)
-    .toLocaleLowerCase()
-    .split('')
-    .map((char) => KARAKALPAK_LATIN_SEARCH_EQUIVALENCE[char] ?? char)
-    .join('');
-}
+const foldKarakalpakLatinForSearch = memoizeLongText(function foldKarakalpakLatinForSearch(value) {
+  return foldKarakalpakChars(normalizeUnicode(value).toLocaleLowerCase());
+});
 
-export function normalizedAliasKeys(value, { transliteration = true } = {}) {
+function computeAliasKeys(value, transliteration) {
   const forms = transliteration
     ? [value, foldCyrillicForSearch(value), foldKazakhForSearch(value), foldKarakalpakLatinForSearch(value)]
     : [value];
-  return [...new Set(forms.flatMap((form) => [
+  return Object.freeze([...new Set(forms.flatMap((form) => [
     normalizeForMatch(form),
     normalizeCompactApostropheForMatch(form),
-  ]).filter(Boolean))];
+  ]).filter(Boolean))]);
+}
+
+// One parse asks every matcher for the keys of the same text, and each
+// derivation folds the text through four scripts. Keyed by text, per mode.
+const aliasKeysWithTransliteration = memoizeLongText((value) => computeAliasKeys(value, true));
+const aliasKeysDirect = memoizeLongText((value) => computeAliasKeys(value, false));
+
+function aliasKeys(value, transliteration) {
+  return (transliteration ? aliasKeysWithTransliteration : aliasKeysDirect)(value);
+}
+
+export function normalizedAliasKeys(value, { transliteration = true } = {}) {
+  return [...aliasKeys(value, transliteration)];
 }
 
 export function aliasesOf(entry) {
@@ -137,7 +167,7 @@ function createAliasIndex(entries, { transliteration = true } = {}) {
   const index = new Map();
   for (const entry of entries || []) {
     for (const value of valuesOf(entry)) {
-      for (const key of normalizedAliasKeys(value, { transliteration })) {
+      for (const key of aliasKeys(value, transliteration)) {
         if (!index.has(key)) index.set(key, entry);
       }
     }
@@ -206,6 +236,32 @@ export function getAliasOwnersIndex(entries, { transliteration = true } = {}) {
   return index;
 }
 
+const PARTIAL_INFO = new WeakMap();
+
+// An alias key is letters and digits joined by single spaces, and so is a text
+// key, so `${alias}` is contained in the padded text exactly when the alias
+// equals a run of whole words of the text. That lets a partial lookup probe the
+// text's word n-grams instead of scanning every alias in the index. `ordinal`
+// keeps each alias's position in the index, which is what orders a tie.
+function partialInfo(index) {
+  let info = PARTIAL_INFO.get(index);
+  if (!info) {
+    const ordinal = new Map();
+    let maxWords = 1;
+    let position = 0;
+    for (const alias of index.keys()) {
+      ordinal.set(alias, position);
+      position += 1;
+      let words = 1;
+      for (let i = alias.indexOf(' '); i !== -1; i = alias.indexOf(' ', i + 1)) words += 1;
+      if (words > maxWords) maxWords = words;
+    }
+    info = { ordinal, maxWords };
+    PARTIAL_INFO.set(index, info);
+  }
+  return info;
+}
+
 /**
  * Every entry tied for the longest partial alias match, in the order they
  * first reach that length. Exact matches always resolve unambiguously to a
@@ -216,25 +272,36 @@ export function getAliasOwnersIndex(entries, { transliteration = true } = {}) {
 export function findCanonicalCandidates(value, entries, { partial = false, transliteration = true } = {}) {
   if (!value) return [];
   const index = getAliasIndex(entries, { transliteration });
-  for (const key of normalizedAliasKeys(value, { transliteration })) {
+  const textKeys = aliasKeys(value, transliteration);
+  for (const key of textKeys) {
     const exact = index.get(key);
     if (exact) return [exact];
   }
   if (!partial) return [];
 
-  const textKeys = normalizedAliasKeys(value, { transliteration });
-  let tied = [];
-  let bestLength = 0;
-  for (const [alias, entry] of index) {
-    if (alias.length < bestLength) continue;
-    if (!textKeys.some((text) => ` ${text} `.includes(` ${alias} `))) continue;
-    if (alias.length > bestLength) {
-      bestLength = alias.length;
-      tied = [entry];
-    } else if (!tied.includes(entry)) {
-      tied.push(entry);
+  const { ordinal, maxWords } = partialInfo(index);
+  const matched = new Map();
+  for (const key of textKeys) {
+    const words = key.split(' ');
+    for (let start = 0; start < words.length; start += 1) {
+      let phrase = '';
+      const stop = Math.min(words.length, start + maxWords);
+      for (let end = start; end < stop; end += 1) {
+        phrase = end === start ? words[end] : `${phrase} ${words[end]}`;
+        const position = ordinal.get(phrase);
+        if (position !== undefined) matched.set(phrase, position);
+      }
     }
   }
+
+  let bestLength = 0;
+  for (const alias of matched.keys()) if (alias.length > bestLength) bestLength = alias.length;
+  const longest = [...matched].filter(([alias]) => alias.length === bestLength).sort((a, b) => a[1] - b[1]);
+  const tied = [];
+  for (const [alias] of longest) {
+    const entry = index.get(alias);
+    if (!tied.includes(entry)) tied.push(entry);
+  }
   return tied;
 }