feat: improve language validation with chunking for description

packages/moderation/src/validation-rules/rules/description.ts

@@ -14,7 +14,7 @@ import {
 import { validateSpam } from '../../validators/spam/index.ts'
 import { evaluateRules } from '../evaluate-rules.ts'
 import {
-	evaluateEnglishText,
+	evaluateEnglishTextBlocks,
 	evaluateNonStandardText,
 	evaluateProfanity,
 	evaluateSlur,
@@ -189,6 +189,16 @@ export function extractDescriptionText(markdown: string): string {
 	return withoutMarkdownSyntax.replace(/\s+/g, ' ').trim()
 }
 
+export function extractDescriptionTextBlocks(markdown: string): string[] {
+	if (!markdown) return []
+
+	return markdown
+		.replace(/```[\s\S]*?```/g, '')
+		.split(/\n\s*\n+/)
+		.map(extractDescriptionText)
+		.filter(Boolean)
+}
+
 export function countText(markdown: string): number {
 	return extractDescriptionText(markdown).length
 }
@@ -248,12 +258,13 @@ export const projectDescriptionValidationRules = {
 	'project-description-non-english': {
 		severity: 'warning',
 		evaluate: (description) => {
-			const text = extractDescriptionText(description ?? '')
+			const blocks = extractDescriptionTextBlocks(description ?? '')
+			const text = blocks.join(' ')
 			if (text.length < MIN_DESCRIPTION_CHARS || !validateSpam(text).valid) {
 				return { valid: true }
 			}
 
-			return evaluateEnglishText(text)
+			return evaluateEnglishTextBlocks(blocks)
 		},
 		presentation: {
 			message: messages.nonEnglish,

packages/moderation/src/validation-rules/text.ts

@@ -1,5 +1,9 @@
 import type { ProjectValidationContext } from '../types/nags.ts'
-import { validateEnglishSummaryText, validateEnglishText } from '../validators/language/index.ts'
+import {
+	validateEnglishSummaryText,
+	validateEnglishText,
+	validateEnglishTextBlocks,
+} from '../validators/language/index.ts'
 import {
 	getNonStandardTextRatio,
 	validateNonStandardText,
@@ -53,6 +57,11 @@ export function evaluateEnglishText(text: string): ValidationRuleEvaluation {
 	return { valid: result.valid }
 }
 
+export function evaluateEnglishTextBlocks(blocks: string[]): ValidationRuleEvaluation {
+	const result = validateEnglishTextBlocks(blocks)
+	return { valid: result.valid }
+}
+
 export function evaluateEnglishSummaryText(text: string): ValidationRuleEvaluation {
 	const result = validateEnglishSummaryText(text)
 	return { valid: result.valid }

packages/moderation/src/validators/language/index.ts

@@ -8,11 +8,26 @@ export interface LanguageDetection {
 export interface EnglishTextResult {
 	valid: boolean
 	detections: LanguageDetection[]
+	reasons: EnglishTextFailureReason[]
+}
+
+export type EnglishTextFailureReason = 'insufficient-english-chunk-coverage'
+
+export interface LanguageChunkAnalysis {
+	totalChunks: number
+	englishChunks: number
+	nonEnglishChunks: number
+	ambiguousChunks: number
+	englishChunkPercentage: number | null
 }
 
 export const MIN_LANGUAGE_DETECTION_WORDS = 8
 export const MIN_LANGUAGE_DETECTION_CHARACTERS = 35
-export const MIN_ENGLISH_SCORE = 0.45
+export const MIN_ENGLISH_SCORE = 0.8
+export const MIN_ENGLISH_SUMMARY_SCORE = 0.5
+export const LANGUAGE_CHUNK_WORDS = 24
+export const LANGUAGE_CHUNK_STRIDE_WORDS = 12
+export const MIN_ENGLISH_CHUNK_PERCENTAGE = 0.3
 
 const wordSegmenter = new Intl.Segmenter(undefined, { granularity: 'word' })
 const characterSegmenter = new Intl.Segmenter(undefined, { granularity: 'grapheme' })
@@ -27,30 +42,104 @@ function hasEnoughCharacters(text: string): boolean {
 	return false
 }
 
-function hasEnoughWords(text: string): boolean {
+function getWords(text: string): string[] {
-	let wordCount = 0
+	return [...wordSegmenter.segment(text)]
-
+		.filter(({ isWordLike }) => isWordLike)
-	for (const { isWordLike } of wordSegmenter.segment(text)) {
+		.map(({ segment }) => segment)
-		if (isWordLike && ++wordCount >= MIN_LANGUAGE_DETECTION_WORDS) return true
-	}
-
-	return false
 }
 
-// TODO, the algorithm can still miss cases where english translation is provided and text primarily in another language
+function getWordWindows(block: string): string[] {
-export function validateEnglishText(text: string): EnglishTextResult {
+	const words = getWords(block)
-	if (!hasEnoughCharacters(text) || !hasEnoughWords(text)) {
+	if (words.length < MIN_LANGUAGE_DETECTION_WORDS) return []
-		return { valid: true, detections: [] }
+	if (words.length <= LANGUAGE_CHUNK_WORDS) {
+		const window = words.join(' ')
+		return hasEnoughCharacters(window) ? [window] : []
 	}
 
-	const results = francAll(text)
+	const starts = new Set<number>()
-	const englishScore = results.find(([language]) => language === 'eng')?.[1] ?? 0
+	for (
-	const detections = results.map(([language, accuracy]) => ({ language, accuracy }))
+		let start = 0;
+		start + MIN_LANGUAGE_DETECTION_WORDS <= words.length;
+		start += LANGUAGE_CHUNK_STRIDE_WORDS
+	) {
+		starts.add(start)
+	}
+	starts.add(words.length - LANGUAGE_CHUNK_WORDS)
+
+	return [...starts]
+		.sort((left, right) => left - right)
+		.map((start) => words.slice(start, start + LANGUAGE_CHUNK_WORDS).join(' '))
+		.filter(hasEnoughCharacters)
+}
+
+export function analyzeLanguageChunks(blocks: string[]): LanguageChunkAnalysis {
+	let englishChunks = 0
+	let nonEnglishChunks = 0
+	let ambiguousChunks = 0
+
+	for (const chunk of blocks.flatMap(getWordWindows)) {
+		const results = francAll(chunk)
+		const primaryLanguage = results[0]?.[0]
+		const englishScore = results.find(([language]) => language === 'eng')?.[1] ?? 0
+
+		if (primaryLanguage === 'eng') englishChunks++
+		else if (primaryLanguage && englishScore < MIN_ENGLISH_SCORE) nonEnglishChunks++
+		else ambiguousChunks++
+	}
+
+	const totalChunks = englishChunks + nonEnglishChunks + ambiguousChunks
+	const classifiedChunks = englishChunks + nonEnglishChunks
+	return {
+		totalChunks,
+		englishChunks,
+		nonEnglishChunks,
+		ambiguousChunks,
+		englishChunkPercentage: classifiedChunks === 0 ? null : englishChunks / classifiedChunks,
+	}
+}
+
+export function validateEnglishTextBlocks(blocks: string[]): EnglishTextResult {
+	const normalizedBlocks = blocks.map((block) => block.trim()).filter(Boolean)
+	const text = normalizedBlocks.join('\n')
+	const chunkAnalysis = analyzeLanguageChunks(normalizedBlocks)
+	const valid =
+		chunkAnalysis.englishChunkPercentage === null ||
+		chunkAnalysis.englishChunkPercentage >= MIN_ENGLISH_CHUNK_PERCENTAGE
+	const detections =
+		getWords(text).length >= MIN_LANGUAGE_DETECTION_WORDS && hasEnoughCharacters(text)
+			? francAll(text).map(([language, accuracy]) => ({ language, accuracy }))
+			: []
 
 	return {
-		valid: englishScore > MIN_ENGLISH_SCORE,
+		valid,
 		detections,
+		reasons: valid ? [] : ['insufficient-english-chunk-coverage'],
 	}
 }
 
-export const validateEnglishSummaryText = validateEnglishText
+export function validateEnglishText(text: string): EnglishTextResult {
+	return validateEnglishTextBlocks(text.split(/\n+/))
+}
+
+export function validateEnglishSummaryText(text: string): EnglishTextResult {
+	const normalizedText = text.trim()
+	if (
+		getWords(normalizedText).length < MIN_LANGUAGE_DETECTION_WORDS ||
+		!hasEnoughCharacters(normalizedText)
+	) {
+		return { valid: true, detections: [], reasons: [] }
+	}
+
+	const detections = francAll(normalizedText).map(([language, accuracy]) => ({
+		language,
+		accuracy,
+	}))
+	const englishScore = detections.find(({ language }) => language === 'eng')?.accuracy ?? 0
+	const valid = englishScore >= MIN_ENGLISH_SUMMARY_SCORE
+
+	return {
+		valid,
+		detections,
+		reasons: valid ? [] : ['insufficient-english-chunk-coverage'],
+	}
+}

packages/moderation/src/validators/language/tests.ts

@@ -4,131 +4,143 @@ import test from 'node:test'
 import { francAll } from 'franc-min'
 
 import {
+	analyzeLanguageChunks,
+	LANGUAGE_CHUNK_STRIDE_WORDS,
+	LANGUAGE_CHUNK_WORDS,
+	MIN_ENGLISH_CHUNK_PERCENTAGE,
 	MIN_ENGLISH_SCORE,
+	MIN_ENGLISH_SUMMARY_SCORE,
 	MIN_LANGUAGE_DETECTION_CHARACTERS,
 	MIN_LANGUAGE_DETECTION_WORDS,
 	validateEnglishSummaryText,
 	validateEnglishText,
+	validateEnglishTextBlocks,
 } from './index.ts'
 
-test('accepts text when franc scores English above the minimum score', () => {
+const english =
-	const text =
+	'This project adds useful tools, configurable settings, and clear documentation for every player.'
-		'This project adds configurable caves, useful tools, and polished world generation for every player.'
+const russian =
-	const result = validateEnglishText(text)
+	'Этот проект добавляет новые инструменты и значительно улучшает игровой процесс для всех игроков.'
-	const english = result.detections.find(({ language }) => language === 'eng')
+
-	const alternative = result.detections.find(({ language }) => language !== 'eng')
+test('accepts English text and retains whole-text language diagnostics', () => {
+	const result = validateEnglishText(english)
+	const englishDetection = result.detections.find(({ language }) => language === 'eng')
 
 	assert.equal(result.valid, true)
-	assert.ok(english)
+	assert.ok(englishDetection)
-	assert.ok(alternative)
+	assert.ok(englishDetection.accuracy > MIN_ENGLISH_SCORE)
-	assert.ok(english.accuracy > MIN_ENGLISH_SCORE)
+	assert.deepEqual(result.reasons, [])
 	assert.deepEqual(
 		result.detections,
-		francAll(text).map(([language, accuracy]) => ({ language, accuracy })),
+		francAll(english).map(([language, accuracy]) => ({ language, accuracy })),
 	)
 })
 
-test('accepts mixed English and Chinese text above the minimum score', () => {
+test('rejects text containing only confidently non-English chunks', () => {
-	const result = validateEnglishText(
-		'A super light QQ bot for minecraft server and QQ group exchange msgs | 超轻量的QQ-MC群服插件',
-	)
-
-	assert.equal(result.valid, true)
-})
-
-test('rejects text when franc scores English below the minimum score', () => {
 	for (const text of [
-		'Чистый модпак для комфортной игры с друзьями, новыми заданиями и значительно улучшенной производительностью.',
+		russian,
 		'これは新しい洞窟と構造物を追加し、すべてのプレイヤーの世界生成を改善するプロジェクトです。',
+		'Um modpack focado em desempenho, imersão e exploração, mantendo a experiência próxima ao jogo original.',
 	]) {
-		assert.equal(validateEnglishText(text).valid, false, text)
+		const result = validateEnglishText(text)
+
+		assert.equal(result.valid, false, text)
+		assert.deepEqual(result.reasons, ['insufficient-english-chunk-coverage'])
 	}
 })
 
-test('validates production project text using the English score threshold', () => {
+test('accepts bilingual text when English chunks are 30% of classified chunks', () => {
-	const cases = [
+	const blocks = [
-		{
+		english,
-			text: 'This mod adds some new things about turtles to Minecraft这个模组为Minecraft增加了一些关于乌龟的新东西',
+		english,
-			valid: true,
+		english,
-		},
+		russian,
-		{
+		russian,
-			text: 'Tenhle Project má super mody ktere zlepší kvalitu hraní PVP',
+		russian,
-			valid: true,
+		russian,
-		},
+		russian,
-		{
+		russian,
-			text: '此插件修复了Authme在lophine服务端上的登录漏洞 修复了玩家退出时SQL数据库的Logged依然为1的问题',
+		russian,
-			valid: false,
-		},
-		{
-			text: 'Мод добавляет рубин — новый драгоценный камень. Добывайте рубиновую руду, кристаллизующуюся в толще камня,  а закалив четыре рубина четырьмя незеритовыми ломами - можно будет сделать меч, кирку, броню и крюк захвата',
-			valid: false,
-		},
-		{
-			text: 'Um modpack Fabric focado em desempenho, imersão e exploração, mantendo a experiência próxima ao Minecraft Vanilla.  O objetivo é melhorar o visual, os sons, a geração de mundo e a qualidade de vida do jogo sem adicionar sistemas complexos.',
-			valid: true,
-		},
-		{
-			text: 'A Create Tacz Warfare Modpack for the Server Create: Warfare',
-			valid: true,
-		},
-		{
-			text: '一个集成了全息字和占位符创建的插件 A Plugin Integrating Holograms and Placeholder Support',
-			valid: true,
-		},
-		{
-			text: "A modpack that adds stuff from TaCZ guns, to shaders, to curios slots, and even create! And also, Superb Warfare, in case TaCZ isn't for you!",
-			valid: true,
-		},
-		{
-			text: 'Leichtes Client-Modpack für entspannte Feierabend-Sessions, bessere Performance, praktische QoL-Mods und ein aufgeräumtes Spielgefühl ohne unnötigen Ballast.',
-			valid: true,
-		},
-		{
-			text: 'A super light QQ bot for minecraft server and QQ group exchange msgs | 超轻量的QQ-MC群服插件',
-			valid: true,
-		},
-		{
-			text: 'Integrates MCP into minecraft, made for mapmakers and complex command block logic and datapack making.',
-			valid: true,
-		},
 	]
+	const analysis = analyzeLanguageChunks(blocks)
+	const result = validateEnglishTextBlocks(blocks)
 
-	for (const { text, valid } of cases) {
+	assert.equal(analysis.englishChunks, 3)
-		assert.equal(validateEnglishText(text).valid, valid, text)
+	assert.equal(analysis.nonEnglishChunks, 7)
-	}
+	assert.equal(analysis.englishChunkPercentage, 0.3)
+	assert.equal(MIN_ENGLISH_CHUNK_PERCENTAGE, 0.3)
+	assert.equal(result.valid, true)
+	assert.deepEqual(result.reasons, [])
 })
 
-test('skips language detection for production text below the word minimum', () => {
+test('rejects mixed-language text when English chunks are below 30% of classified chunks', () => {
-	for (const text of [
+	const blocks = [english, english, russian, russian, russian, russian, russian]
-		'You can chat Gemini AI in Minecraft',
+	const analysis = analyzeLanguageChunks(blocks)
-		'Create Mods X Zombie Apolcalypse',
+	const result = validateEnglishTextBlocks(blocks)
-		'Open-world zombie survival modpack',
-		"BIG-GOOSE Minecraft server's modpack",
-	]) {
-		assert.deepEqual(validateEnglishText(text), { valid: true, detections: [] }, text)
-	}
-})
-
-test('skips language detection for text below the character minimum', () => {
-	const result = validateEnglishText('one two three four a b c d')
-
-	assert.deepEqual(result, { valid: true, detections: [] })
-	assert.equal(MIN_LANGUAGE_DETECTION_CHARACTERS, 35)
-})
-
-test('validates production text meeting both signal minimums', () => {
-	const text = '𝗔𝗶𝗺𝗶𝗻𝗴 𝘁𝗼 𝗲𝗻𝗵𝗮𝗻𝗰𝗲 𝗠𝗶𝗻𝗲𝗰𝗿𝗮𝗳𝘁 𝘄𝗵𝗶𝗹𝗲 𝗿𝗲𝘁𝗮𝗶𝗻𝗶𝗻𝗴 𝘁𝗵𝗮𝘁 𝗩𝗮𝗻𝗶𝗹𝗹𝗮 𝗳𝗲𝗲𝗹!'
-	const result = validateEnglishText(text)
 
+	assert.equal(analysis.englishChunks, 2)
+	assert.equal(analysis.nonEnglishChunks, 5)
+	assert.ok((analysis.englishChunkPercentage ?? 0) < MIN_ENGLISH_CHUNK_PERCENTAGE)
 	assert.equal(result.valid, false)
-	assert.ok(result.detections.length > 0)
+	assert.deepEqual(result.reasons, ['insufficient-english-chunk-coverage'])
+})
+
+test('uses overlapping 24-word windows with a 12-word stride', () => {
+	const text = Array.from({ length: 48 }, (_, index) => `word${index}`).join(' ')
+	const analysis = analyzeLanguageChunks([text])
+
+	assert.equal(LANGUAGE_CHUNK_WORDS, 24)
+	assert.equal(LANGUAGE_CHUNK_STRIDE_WORDS, 12)
+	assert.equal(analysis.totalChunks, 4)
+})
+
+test('skips chunks below the minimum word count', () => {
+	for (const text of [
+		'Minecraft',
+		'Minecraft server',
+		'This description has only seven English words',
+	]) {
+		assert.deepEqual(validateEnglishText(text), { valid: true, detections: [], reasons: [] })
+	}
+
 	assert.equal(MIN_LANGUAGE_DETECTION_WORDS, 8)
 })
 
-test('uses the same validation for summaries', () => {
+test('skips chunks below the minimum character count', () => {
-	assert.equal(validateEnglishSummaryText, validateEnglishText)
+	assert.deepEqual(validateEnglishText('a b c d e f g h'), {
+		valid: true,
+		detections: [],
+		reasons: [],
+	})
+	assert.equal(MIN_LANGUAGE_DETECTION_CHARACTERS, 35)
+})
+
+test('accepts summaries with an English score of at least 50% without changing description logic', () => {
+	const text =
+		'Um modpack Fabric focado em desempenho, imersão e exploração, mantendo a experiência próxima ao Minecraft Vanilla.'
+	const summaryResult = validateEnglishSummaryText(text)
+	const englishDetection = summaryResult.detections.find(({ language }) => language === 'eng')
+
+	assert.equal(MIN_ENGLISH_SUMMARY_SCORE, 0.5)
+	assert.ok(englishDetection)
+	assert.ok(englishDetection.accuracy >= MIN_ENGLISH_SUMMARY_SCORE)
+	assert.equal(summaryResult.valid, true)
+	assert.equal(validateEnglishText(text).valid, false)
+})
+
+test('skips summary detection below its word or character minimum', () => {
+	for (const text of ['This summary has only seven English words', 'a b c d e f g h']) {
+		assert.deepEqual(validateEnglishSummaryText(text), {
+			valid: true,
+			detections: [],
+			reasons: [],
+		})
+	}
+
+	assert.equal(MIN_LANGUAGE_DETECTION_WORDS, 8)
+	assert.equal(MIN_LANGUAGE_DETECTION_CHARACTERS, 35)
 })
 
 test('allows empty text to be handled by required-field validation', () => {
-	assert.deepEqual(validateEnglishText('  '), { valid: true, detections: [] })
+	assert.deepEqual(validateEnglishText('  '), { valid: true, detections: [], reasons: [] })
 })