How the Korean glossary is built

The two sides

• English. The corpus on this site, every section checked word for word against the file bahai.org served.
• Korean. The official translations published at bahai.or.kr.
• Neither side is edited, corrected or retranslated. The pipeline only reads them.

Step one: pair the passages

Nothing can be learned about a word until the glossary knows which Korean paragraph translates which English one. A wrong pairing does not fail loudly. It quietly invents a translation no publisher printed, and the invented term then looks as well attested as a real one. So every book is keyed on the strongest thing the two editions actually share, and four kinds of book turned up.

1. Both editions print the same number (10 books, 986 passages)

• Pair on the printed number. It is a fact each publisher recorded, so the pairing is proven rather than estimated.
• The number differs by book: a paragraph number in the Kitáb-i-Aqdas, a selection number in Gleanings, a verse number in The Hidden Words, a chapter in Some Answered Questions.
• Every number must find exactly one counterpart. Anything unmatched is reported, never guessed at, so the Korean Questions and Answers stopping at 95 of the English 107 shows up as twelve missing rather than twelve approximations.

2. They share a number, but not an obvious one

• Paris Talks. The English numbers its talks 1 to 64; the Korean gives only titles. Both record the date each talk was given, and a date means the same thing in either language, so the date is the key.
• Two wrinkles there. The Korean date lines carry a weekday and often the address, so the first month and day is taken rather than requiring the whole line to be a date. And two talks can fall on one day, so the key records which occurrence it is.
• The Bishárát. The Korean numbers its own paragraphs 1 to 31 on a scheme the English does not share, which looks like a key and is not. What both print is the ordinal announcing each Glad-Tidings: "The third Glad-Tidings" against 세번째 기쁜 소식. Fifteen on each side, all fifteen aligned, nothing unmatched.
• Obligatory Prayer and Fasting. Three numbered sections on both sides, with Roman numerals that restart inside each, so the key is the section and the numeral together.
• That book is 53 of 54, and the miss is a misprint rather than a fault. The Korean page numbers an extract XXXIV in a run of twenty-seven, between XXIII and XXV, where the sequence needs XXIV. Which extract it is, is obvious to a reader. The pairing still refuses to guess it and reports both sides unmatched.

3. Neither edition numbers anything (6 books, 1,426 passages)

• The Kitáb-i-Íqán prints no numbers in either edition, and its footnotes are no help: 186 notes in English against 257 in Korean, on different schemes.
• What both editions keep is order. A Korean paragraph runs about half the length of the English it translates, and that ratio holds steadily through a book, so the two texts are walked down together with the lengths kept in proportion. The ratio is measured from each book rather than assumed: 1.96 in Part One of the Íqán, 1.87 in Part Two.
• Length alone cannot see a slip of one paragraph, because every pairing after the slip is still about the right length. So a candidate pairing also costs the share of the terms in its English that its Korean does not show, counted on terminology the other books have established, with half the lexicon held back to score the result. That found and fixed a slip in the longest House of Justice message, which is below.
• Where a shared division exists it bounds the pairing, so drift cannot cross it. The Promise of World Peace has a preamble and four numbered sections on both sides. The Seven Valleys has the valleys, which the Korean tags and the English announces in words.
• The Unity to Contentment transition is the one the English does not announce, so those two valleys are read as a single stretch rather than split at a guessed point.
• Twenty messages of the Universal House of Justice are the one case where the pairing is not deduced at all. The Korean site publishes each of them in both languages, as two pages named alike but for an e or a k. That naming is the publisher saying which two pages carry the same message, which beats anything the texts could be asked to show for themselves. Each message is then aligned on its own, so drift in one cannot reach the next.
• The other Korean message pages are published in Korean only. Their English is the message as bahai.org publishes it, which is now in this library, and each Korean page is paired with the message bearing the date the page is named for: a page named for a year alone is that year's Riḍván message. Pages from 2021 on that hold the English with no Korean at all are left out, as are those whose message bahai.org does not list.
• For those twenty the Korean site's English page settles only which message it is. The English read against the Korean is bahai.org's own, the same as for every other message, so a citation points at the message in this library.
• Reading them is worth the trouble because they are modern letters rather than scripture. The nineteenth-century works, however many are added, keep returning the same divine names and attributes; the messages bring vocabulary none of them contain.
• Editions break paragraphs differently, and the differences are handled structurally rather than absorbed. The Korean Seven Valleys sets poetry in its own paragraphs where the English keeps it inside the numbered one, so a verse is folded back. The Korean also carries translator's notes with no English counterpart, so those are dropped.
• Front matter is not prose and has nothing to pair with. The English Promise prints the signatory, date and addressee as three short lines where the Korean puts two of them in a heading, so a minimum length leaves them out of both sides.

4. One edition carries only a selection (1 book, 105 passages)

• The Korean Prayer and Devotional Life prints 105 extracts against the English 135, and its meditation section carries 6 where the English has 20.
• Order is no help. Keeping the two in proportion pushed three English extracts onto one Korean, and reading that pair showed the pooled English was about meditation granting eternal life while the Korean extract was about the sign of intelligence being contemplation. Different passages entirely.
• A selection says nothing about proportions, so the question is not how the passages were broken up but which one each Korean extract translates. Extracts are matched one by one on the terminology they share, and the 30 the Korean edition did not carry are left unpaired.

Leaving a passage unpaired

• An edition may simply not carry a passage. Without a way to pass one over, the pairing has to fold it into its neighbour, which invents a pairing rather than reporting a gap.
• So a passage can be skipped, priced between a genuine merge and a bad one. The Íqán and the Promise take no skips at all; the Seven Valleys takes one.

Step two: find the Korean term

• Each English term is tested by contrast. The paired passages fall into two groups: those whose English contains the term and those that do not.
• The Korean for that term is the run of characters that keeps turning up in the first group and stays out of the second.
• No Korean dictionary, word list or language model takes part. The Korean is identified only by where it does and does not appear, which is why every entry traces back to the translations rather than to an outside source.
• Korean attaches particles to the end of a word, so a winning run can carry one. A particle is only stripped where the corpus shows the remainder to be a word in its own right, meaning it is seen taking other endings and seen bare.
• The run that matches best is often the part several written forms share, so it stops short of a whole word. All-Merciful matches 가장 자비 where the translations write 가장 자비로우신, and All-Glorious matches 가장 영광, which appears nowhere on its own.
• So the headword is the most frequent form actually written, and the matched run stays on the entry as the thing the confidence describes. Every written form is listed with its count, so 모함마드 as a headword still shows 무함마드 beneath it.
• Transliteration is inconsistent even within the official translations, so one term can have several renderings. All are kept with their counts and the most frequent becomes the headword.

Names are heard, not read

• Contrast alone is deaf to a name. Korean writes a foreign name by its sound, and the test only knows which passages a string appears in, so a phrase that keeps a name company can out-cover the name itself. Shoghi Effendi resolved to 같이 언급하셨습, "commented as follows", which appeared in three of his four passages where 쇼기 에펜디 appeared in two.
• So the Hangul is romanised and stripped to its consonants, which is where two transliterations of one name agree even when their vowels do not: 모함마드 and 무함마드 are both mhmmd, and so is Muhammad.
• Length has to agree as well as letters, or a phrase reaches the threshold by opening on the right sounds. Germany against 그러므로 우리는, "therefore we", matches g, r, m and n inside a string half as long again.
• The resemblance is a gate rather than a nudge. Allowed to tilt the choice by degrees it slid Christ from 그리스도 to 그리스, which is the Korean for Greece, so a partial resemblance now counts for nothing.
• A near-exact match is evidence in its own right, so a name can be published on it where coverage alone would never carry it. 쇼기 에펜디 names him in two of four passages, and it is plainly his name. 24 entries stand on that, most of them places and people.

What confidence means

• It balances two failures: passages where the English term appears and the Korean form does not, and passages where the Korean form appears without the English term.
• It is tempered by how much evidence stands behind it, so a term seen three times cannot score as though it were seen three hundred.
• A middling score usually means one of two things, worth telling apart: the term has more than one Korean rendering, or the Korean phrasing is looser than a single run of characters can capture.
• An entry scoring below 0.55 is not published. That bar used to be applied only to an intermediate score, so entries reporting as low as 0.03 were reaching the glossary: Mover resolved to 제가 across 54 passages that had nothing to do with it. Enforcing the bar on the reported score cut the glossary by about two thirds.
• A single word of a phrase, resolving to the phrase's Korean, is not a term of its own. Candidates include capitalised words as well as phrases, so All-Glorious also produced Glorious and Maids of Heaven also produced Maids. Entering both states something false, that Glorious alone means 가장 영광되신, so the word is dropped and the phrase kept.
• Where two terms resolve to the same Korean and neither contains the other, both stand and each is flagged against the other. That is the method's blind spot rather than a duplicate: two words appearing in the same passages cannot be separated by which passages they appear in, so East resolves to 서양, which is West.
• 62 of the 224 entries score 0.75 or above.

The strongest entries at present:

Checking a pairing that no number proves

• Blind test. Take a book that does print numbers, hide them, run ten adjacent Korean paragraphs together the way the Korean Íqán does, and check the pairing puts them back. Across four books it recovers at least 97 percent of numbering it was never shown.
• Terminology check. Score every resulting pair against the glossary deduced from the other books, holding out any term only this book evidences so it cannot mark its own homework.
• A control, always. The same pairings shifted by one passage are scored too. The mean alone says nothing, because how much of the glossary a book uses is a fact about the book. The gap between mean and control is the reading.

Read that table as a guard against gross misalignment, not as a ranking. The check needs passages long enough to carry distinct terminology, and the two shortest books in the corpus score worst. Aqdas, Synopsis averages 65 characters a passage and scores 0.71 against a control of 0.14, below its own control, even though it is paired on a number both editions print and its pairings are therefore certain. Where neighbouring passages are that short and that alike, shifting by one barely changes the vocabulary. A weak score there means the check is blind, not that the pairing is wrong.

Where the translations disagree

• Reducing a term to one headword throws something away. Two Korean forms for one English term is also a fact about the translations, so 14 terms written more than one way are kept as a record.
• Korean grammar is not a disagreement. A form inside another is a plural or a title, and a stem taking two endings is a conjugation, so neither is counted.
• Telling that apart from real variance is harder than it sounds: 아름다움 against 아름다운 and 페르시아 against 페르샤 both differ by one letter near the end. What separates them is where the difference falls and how long the endings are.
• Co-occurrence is not a disagreement either. 바하올라 scores 0.60 against "Question", which is Bahá'u'lláh keeping company with the word rather than translating it, so a form another term claims more strongly is dropped.
• The labels below sort spelling from word choice, and they get it wrong often enough to say so: 페르시아 and 페르샤 are one name filed under a different word. Read them as a rough sort.

The books each form appears in are the useful part. 페르샤 for Persia turns up only in Paris Talks. Ordainer is 제정자 in Gleanings and 결정자 in the Kitáb-i-Aqdas, with neither book using the other's word. A record like that gets more useful as books are added, not less.

What went wrong on the way

Each of these produced plausible output for a while, which is the point of writing them down.

• Length alone paired text with the wrong paragraph. Around Part One paragraph 43 of the Íqán, one Korean paragraph carries the end of one English paragraph and the start of the next, so the boundaries cross rather than merge. Without a two-against-two step the pairing faked the crossing with a one-against-two and a two-against-one, balanced the lengths perfectly, and pinned Korean text to the paragraph beside the one it belonged to. Two wrong pairs that looked right on every number available.
• Particle stripping invented words. Christ came out as 그리스, which is the Korean for Greece. Mary came out as 마리, and Son of Man as 사람의 아.
• The first fix for that was also wrong. Requiring the shortened form to appear in more passages proves nothing, because a shorter string sits inside at least as many passages as a longer one, so almost any truncation passes. What works is asking what follows the remainder: 하느님 is followed by 의, 을, 께 and by nothing at all, while 마리 is followed by 아 every single time.
• The same fault hid in a second place. Persia was recorded as 페르시 when the translations write 페르시아, because the final 아 is a particle elsewhere and was stripped on sight.
• A shared core swallowed the spellings it stood for. 함마드 sits inside both 모함마드 and 무함마드, so counting it as a third way of writing the name left Muḥammad with one form and no disagreement to report.
• Sentence openers became terms. The filter for grammatical scaffolding was applied only to multi-word candidates, so The, And, Now and This came through as terms of their own. Nothing in Korean answers to any of them, so each was an entry that could never be satisfied.
• An English message came out shuffled into nonsense. The December 2021 message read "These three constant potential on its own", two halves of different sentences stuck together. The page was fine. The reader had broken each paragraph at the newlines the source file indents its markup with, then sorted the fragments by where they started, and since every fragment of one paragraph starts at the same place the tie was settled alphabetically. Only a real line break divides a block now, and fragments keep the order they were written in.
• The check is what caught it. That message scored 0.06 above its shifted control when every other message scored between 0.13 and 0.70. Nothing else pointed at it, and the text is long enough that nobody was going to read it through. After the fix it scores 0.37.
• That last one hid the quality of everything else. The Seven Valleys looked like the weakest book in the corpus, scoring 0.37 against a 0.20 control. What settled it was scoring only the seven valley openings that had been checked by reading: those scored 0.43, which said the instrument was wrong rather than the book. With the pseudo-terms removed the book reads 0.72, and all three books paired by order sit inside the range the books proven by number occupy.
• Every entry appeared to come from Gleanings. The passages cited were the first few in corpus order, and Gleanings is the first book the pipeline lists. References are now sampled one per book before a second from any.
• A filter read a sample as though it were a fact. Whether an entry counted as proven depended on which eight references happened to be kept, so each entry now records every book it was found in.
• One book was never a translation of the book beside it. The Kitáb-i-Aqdas Questions and Answers was paired against a Korean file of 95 numbered questions. That file is an introductory catechism, opening "is there a purpose to our living?" and going on to "who is Bahá'u'lláh?". The Aqdas answers questions on inheritance, ablutions and prostration, and none of that vocabulary is in the Korean at all. Both sides number their items, 107 against 95, so pairing by number produced 95 confident pairs of unrelated text.
• No rule was ever going to catch that, and the rule worked. Pair two numbered lists by number is correct, and it ran correctly. What was missing is the question a reader asks first and no measurement asks at all: are these two documents the same work? The check did flag the book, with the one negative separation in the corpus, meaning its pairings read worse than deliberately wrong ones. That was read as a weak book rather than a wrong one for as long as it sat there. Removing it took the corpus from 1,603 pairs to 1,508 and the glossary from 212 entries to 224, dropping none: the bad pairs were not inventing entries, they were diluting the contrast and suppressing twelve real ones.
• So the check now prints the text. It shows one passage from the middle of every book, English above Korean, before any of the figures. There are thirteen books. Reading thirteen pairs of passages takes a minute and settles in a glance what no amount of scoring can reach.
• The sound test scored every short name 1.00, including the wrong ones. Matching a name against its Korean drops the vowels, because that is where two transliterations of one name disagree. For a short name that leaves almost nothing: Haifa keeps p alone and Bahai keeps b. One letter matches one letter perfectly, so 아 바 scored exactly what 하이파 would, and anything with two passages behind it could be published on sound alone. That is how Bahai was published as 아 바 and Chile as 호꾸꼴, which is part of Ḥuqúqu'lláh.
• The same blindness stopped names being trimmed. The step that drops the words around a name compares sounds, and where every form scores 1.00 it has nothing to compare, so Salsabíl kept the particle after it as 살사빌을. Four consonants are now treated as enough to mean something and a shorter match is worth proportionately less, so a short name has to be carried part of the way by the passages it appears in. Three entries went; the twenty-one names that were right stayed.
• The prayer book defeated the method, and is not in the corpus. The Korean prints 184 prayers and the English 270, in different orders, numbered on neither side. Matching them on the glossary got about a third right. The reason is the glossary itself: it is largely divine names and attributes, and every prayer carries them, a mean of 5.1 known terms and never zero. Vocabulary that appears everywhere cannot tell one prayer from another.
• Its topic headings could not be mapped either. 34 English topics against 42 Korean sections, each alphabetical in its own language. Reading them by sound and by glossary matched 6, and put Morning against 망자, the departed, at a higher score than it gave the two that are plainly right, Naw-Rúz against 노우루즈 and Ḥuqúqu'lláh against 호꾸꼴라. Strings that short cannot be told apart this way. Pairing the prayers by hand would mean deciding what the Korean means, which is the one thing this method does not do.

Where the messages pairing slipped, and what fixed it

The messages of the Universal House of Justice were then not one book but twenty documents, and neither edition numbers its paragraphs, so they are paired by order within each message. That is an estimate. Scoring the book as a whole said it was healthy, and the book was the wrong unit: a single long document can slide while the average stays comfortable. Scoring each message on its own found two that were not healthy, the two longest, which is the signature of a pairing that is right at the start and drifts later.

Scoring each pairing within one message found exactly where. In the message of 28 November 2023 the first fifty pairings score between 0.85 and 1.00 and their neighbours score far lower, so they are right. From the fifty-first on, the score falls to around 0.3 and the neighbouring Korean starts scoring higher than the paired Korean, which is not weak evidence of a pairing but good evidence of the wrong one. Everything after the slip was off by a paragraph, and nothing in the length figures could show it, because a pairing off by one is still about the right length. Only the words change.

So the words are now priced into the alignment. A candidate pairing costs what it deviates from the book’s own length ratio, as before, plus the share of the terms in the English that the Korean does not show, counted on the terminology the other books have already established. The Korean edition of that message is 72 paragraphs against 88 English and carries 93% of the text the book’s ratio predicts, so it is merging paragraphs rather than dropping them, and the alignment needed to find the merges in the right places rather than to be told how many there are.

Terminology in the cost would otherwise mark its own homework, since a pairing built to make known terms line up will of course show known terms lining up. So the lexicon is split in two by a hash of the term: the alignment is given one half and scored on the other, which it never sees. On those held-out terms the message of 28 November 2023 goes from 0.855 to 0.945, and the pairings its own neighbour outscores go from nine to two. The message of 2010 goes from 0.654 to 0.917 and from five to none. Every message now scores 0.77 or better, the weakest being the message of 2008 at 0.769 against a control of 0.312, and the book as a whole scores 0.937 against 0.563. The Kitáb-i-Íqán and The Promise of World Peace do not move, and the Seven Valleys gains a little, so the same weight is used for all four books paired this way.

Three terms came back with the fix. The Guardian’s, the Bahá’ís of Iran and the Bahá’ís of the World were all recorded as carrying nothing in Korean, and all three were only unreadable because their passages sat after the slip. The Korean has 수호자, 이란 바하이 and 세계 바하이 for them. That leaves four terms with nothing in Korean, and all four are the known artefact: a capitalised run that ends on an adjective because the next word is lowercase, as in the Dayspring of Divine.

korean/tools/messages.py prints the paragraph counts per message, and the per-message and per-pairing scores can be re-run whenever a message is added.

Checking the readings against the text

The readings are the only thing asserted here, and reading is what produced them, so a mistake made while reading cannot be caught by reading again. It also hides itself. A form written slightly wrong matches nothing, and a form that matches nothing looks exactly like a form nobody has recorded: the term shows a gap, the reading is never printed next to a passage, and so nothing ever contradicts it. The check puts every recorded form back against the Korean of the passages its own term is credited with.

Forty-nine forms reached nothing, in four kinds.

• The stem was not where the word ends. Korean adds its grammar on the end, so a form is recorded as the run to search for, and a run that stops mid-word matches nothing: 요단 is only ever written 요단강, 감추어 only 감추어져, 능력이 있 only 능력이 있사옵나이다.
• The brackets and the spacing were not what the text has. The translations put brackets round a name inside a phrase and write the Counsellors as one word, so the forms are 「에덴」동산, 계시의 거룩한 「붓」 and 대륙고문, not the tidier versions that were recorded.
• The form was read off a passage that belongs to a longer term. 대군중 was read off the Concourse on high and left on Concourse, 지고하신 off the Pen of the Most High and left on Most High. The longer term had them recorded already, so the copies said nothing their own passages say.
• One was a plain misreading. Beloved carried 사랑하는 이, read out of 사랑하는 이란의, beloved Iran’s, where the 이 is the first syllable of Iran and belongs to it.

Two of the corrections are worth having for themselves. 모세오경 is the Korean for the Pentateuch and not for the Torah, and Gleanings uses it where the English says Pentateuch. And Gleanings renders the Essences of Detachment 초연의 본질, aloofness, where the Kitáb-i-Íqán has 무집착의 본질, non-attachment.

The first kind of fault also had a cause worth fixing rather than working around. The translations gloss a Korean word by putting its hanja in brackets straight after it, so the Íqán writes the Essence of Essences as 정수(精髓) 중의 정수. The gloss annotates the word rather than belonging to it, and because it sits inside the phrase it breaks any form longer than one word. There are 1,797 of them in 608 passages. They are now levelled out for matching, and only brackets holding nothing but ideographs and spaces are touched, so a footnote marker like (주1) and anything written in Korean are left alone. What a gloss says is never consulted, which keeps this orthography and not meaning.

One addition to the grammar came out of the same check. Korean contracts the copula onto the end of a word to turn it into a modifier of the next one, as in 지휘자인 알리, Ali who is the Commander, and that ending was missing from the list of things that may follow a word without being more of it. Adding it reached twenty-seven more passages.

korean/tools/forms.py runs the check and now reports none, of the 2,107 forms recorded. Sixty-three more are redundant rather than wrong: they do match, and another form of the same term matches a longer run in the same passage, so only the longer one is ever shown.

The same English, asked twice

Gleanings is a compilation, and part of what it compiles is published in full elsewhere in this corpus. Forty-one of its selections are passages of the Kitáb-i-Aqdas and thirty-five are passages of the Kitáb-i-Íqán, and the English is word for word the same, because one translator made all three: Shoghi Effendi. That makes a test nothing else here performs. The English side asks the Korean editions the same question twice, in the same words, and the answers can simply be compared.

They mostly differ. Of 479 terms that appear in both halves of a shared passage with Korean recorded on both sides, 200 are written one way in one book and another way in the other, and 279 agree. The Lord alone accounts for nine of the disagreements, the Beloved for six, and the King, Muḥammad, the Presence and the All-Glorious for five each. These are not two shades of meaning in two different sentences, which is what a disagreement between two books usually is. It is one sentence, translated twice, and the two Korean editions of it do not match.

Matching the passages is done on word runs rather than on equality, because a compilation trims and joins: a passage counts as shared when the two hold a run of ten words in common and the shorter is mostly inside the longer. Coverage is measured both ways as well, because a Gleanings selection and an Íqán paragraph often hold the same sentences without holding the same text, one starting where the other is half over. The terms can be compared in either case, since both passages carry the term in English. The two Korean texts can only be compared with each other where neither passage is mostly something the other does not hold, and five of the eighty-nine are like that. There the two translations have 16% of their five-character runs in common, and the closest pair, a Hidden Word reprinted in Prayer and Devotional Life, has 29%.

python -m korean.kw.cli echoes writes the comparison, and the catalogue shows the worst of it above the table.

Where a Korean word ends

Everything above turns on one test: given a run of Korean recorded for a term, does this passage use it as a word or only contain its syllables. Korean writes no spaces inside a word and adds its grammar on the end, so a loose test counts 영 for Spirit inside 영광, glory, and 영원, eternal. The test is that nothing of a word comes before the run, and what follows it is either not a Korean syllable at all or one of the endings the language attaches. That is grammar, and it says nothing about what any word means.

Checking every recorded form against its own passages showed where that test was too strict and where it was too loose. Four changes came out of it, each measured on the passages it moved.

• Spacing is not fixed. One book writes 쇼기 에펜디 and another 쇼기에펜디, one message writes 신성한 계획 and 신성한계획, and a stray space turns up inside 하느님 의 where the particle belongs to the word. So a form tolerates whitespace between its characters, which reached 26 more passages. Two syllables are too short for it, because Korean marks grammar with one syllable on the end of a word: 의사 for Physician was being found in 의의 사원에, the temple of justice, and 가지 for Branch in 이사야가 지적한, Isaiah pointed out.
• A one-syllable form is held to a particle. 주 is the Lord and also the stem of 주다, to give, and the passages write both 주께서, the Lord, and 주시는, giving. One syllable cannot tell them apart, and what follows a noun is a case marker, so a form that short is allowed only those. Recording 주 then took the Lord from 62 passages uncovered to 7.
• A three-syllable word may carry a compound. Korean writes a compound without a space, so the strict test refused 바하이공동체 as a use of 바하이, 시나이산 as a use of 시나이 and 불란서사람 as a use of 불란서. Three syllables is where the compound becomes the likelier reading, since nothing but the term itself opens 바하이. It is the form’s last word that has to be long enough, so 거룩한 영 stays strict and is not credited with 거룩한 영감, a holy inspiration. That reached 35 more passages.
• The words shown now obey the same test as the words counted. A form is often recorded as a stem, because that is what several written forms share, and the catalogue shows the whole words from the text instead. Those were collected without the boundary test, so the spellings printed under a form were not always spellings the form had matched: 영 for Spirit was listed as 영광의 and 영원한, 펜 for Pen as 펜디의 out of Shoghi Effendi, and 때 for Time as 때문입니다, because. The counts beside them were wrong in the same way.

One syllable is still one syllable. 새 for Bird is credited in 새로운, new, because 새로 is 그 word with an adverbial marker on it and nothing short of knowing what the sentence says can separate the two. The catalogue shows the word it matched, so a reader can see it.

A second method, and what the first one’s rules did

Everything above answers one question: which Korean form is the equivalent of this English term. That question needs a confidence score, and a score drops the terms whose Korean is least consistent, which are the ones most worth recording. So there is now a second page, the catalogue, which asks a different question: what do the official translations write for this term. It lists every form with its count, and a term may carry four answers if that is what the books do. It holds 1,610 terms, with the Korean read for 1,610 of them so far and 3,576 forms recorded.

Whether a term is in the catalogue is decided on the English alone, so no Korean evidence adds or removes one. The English side is read for capitalised words and phrases, sentence grammar is taken out of what comes back, and the rest is kept. Capitalisation is doing two different jobs in scripture, which is the whole difficulty: it marks the terms, the Cause, the Manifestation, the Ancient Beauty, and it also marks the first word of every sentence, so Verily, Likewise and Whoso arrive looking exactly like terminology.

Sorting that out took a list of rules, and the rules were wrong in ways that only showed up in the output. These are worth setting down, because none of them was a mistake about Korean or about what a term is. Each was a rule that was right about the case it was written for and wrong about the next one.

• One list was doing two jobs. A word was listed as grammar both to keep it out of the catalogue as a term of its own and to trim it off the end of a phrase. Glory was listed for the sake of “Glory be to God”, and that cut the Dayspring of Glory, the Pen of Glory and the Kingdom of Glory back to their first word, below the two-word floor, where they disappeared. House was listed for the Paris Talks headings, which give the address each talk was given at, and it took the House of Justice and the Houses of Worship with it. Most was on the list twice, once as grammar and once as a modifier, and the grammar won, so the Most Great Prison came out as Prison and the Most Holy Book as Book. Every Most term on the catalogue page had been entered by hand. Separating the two jobs, so that a function word is trimmed and a noun never is, recovered 70 terms, Most High among them, which has 43 passages of its own.
• A whole shape of term could not be seen. The pattern allowed one small word between two capitalised ones, and “of the” is two, so anything shaped X of the Y was invisible: the Pen of the Most High at 24 passages, the Seal of the Prophets, the Shrine of the Báb, the Hands of the Cause, the Point of the Bayán, the Commander of the Faithful. Allowing a preposition to bring its article with it found 48 more terms, and letting a phrase inside a longer phrase be counted and then kept only if the text somewhere writes it on its own, which is how the Mosaic Dispensation survives the Law of the Mosaic Dispensation, found 44 more.
• The alphabet was too small. The letters a term may hold were written as the Latin range that covers á and í, which leaves out the underdotted ḥ, ḍ, ṣ, ṭ and ẓ the transliteration also uses. So Muḥammad was cut in two at the ḥ, and the corpus held a term called Mu and another called ammad, while Imáms of the Muḥammadan Faith was catalogued as Imáms of the Mu. Riḍván, which names both the garden and the twelve days of the festival, was missing altogether, 35 passages of it.
• One name appeared twice as two people. The corpus writes ‘Abdu’l-Bahá with an ordinary hyphen in some places and a non-breaking hyphen in others. They are two characters and one name, so the catalogue carried two rows, 48 passages and 16.
• The tool for reading the Korean showed the wrong passages. The catalogue is careful not to credit a short term with a longer term’s Korean, so Justice is not given 정의원, which is the House of Justice. The tool that puts the passages in front of a reader was not doing the same, so Days was shown 34 passages of which 28 were Ancient of Days. Reading those would have filed 태고부터 계신 분 under Days. The two now use one test.

The common thread is that each rule was reasonable on its own and the pile of them was not. A person looking at a list of Bahá’í terms with no House of Justice, no Most Great Prison and no Riḍván in it would see the problem at once. Nothing in the figures showed it, because the figures were about the Korean and the fault was in the English.

What counting alone can find

The reading is finished: all 18,040 places where a term in the English meets a Korean passage are either covered by a recorded form or recorded by hand as carrying nothing for the term. That leaves a different question, and a useful one for a translator. Some faults in a text can be found without reading it at all, by counting one thing against another, and a count can go through every passage without getting tired. Five of these run over the Korean side and feed the list for a translator.

• Brackets and speech marks that do not close. The openings are counted against the closings in each passage, which finds 107 places in 95 of them. Footnote numbering is left out, since a digit followed by a lone closing bracket is how these editions number a note, and so are the single quotation marks, because the right single quote is also the apostrophe in Ka’bih.
• Particles standing as words. Korean writes its grammar on the end of the word it belongs to, so a particle with a space in front of it is a space in the wrong place: 344 of them, and 154 are in the Seven Valleys. Only the particles that cannot be a word in their own right are counted.
• Hanja glosses longer than their word. A gloss repeats the word in hanja one character to a syllable, so counting the characters against the syllables finds 8 that claim more than the hangul in front of them has. One of them turned out to be the wrong character: anthropomorphism is glossed 神人同型同性設, ending in the character for establishing where the word is the one for a doctrine.
• Counted numbers that differ. A number with its noun behind it in the English against a numeral with its counter behind it in the Korean, which is the only way to compare them, since the English spells its numbers out and the Korean writes digits or Sino-Korean numerals. A single numeral syllable is not read as a number, because 이 is also this and 만 is also only. Thirty-seven passages came out of it and one is a plain error: the earth completes a revolution every 365 days, five hours and forty-eight minutes in the English, and 56 hours in Some Answered Questions 41.
• Latin letters inside the Korean. Almost all of these are the transliteration the editions print beside a name, 카프Káf and 자카트Zakát. Two were not: the Korean Íqán source writes a footnote tag without its opening bracket, at footnotes 115 and 132, so the words sup class="fn" sat in the middle of two sentences. The bracket is put back before the tags are stripped.

A sixth count reads the forms rather than the text. Korean marks respect for the subject with 시 inside the verb, where it contracts with what follows, so 전능하신 and 전능한 are the same word spoken two ways and do not look alike. 29 terms are written both ways, and for a name of God the question is not which word the book chooses but how it speaks to God. The Aqdas makes the choice both ways in one book: 전능하신 four times, 전능한 eight.

One more count turns the catalogue over rather than reading the text. The catalogue asks what Korean an English term is given; asking which English terms share one Korean form puts the question the way a Korean reader is in it, and 173 forms carry 382 terms between them once terms that are versions of each other are left out. 하느님 is God and also Lord, 경전 is the Book and also the Scriptures and the Writings, 태양 is the Sun, the Daystar, the Luminary and the Orb. None of that is a fault. It is the distinction the English makes and the Korean does not, which is the thing a reviser most needs to see.

Two counts found nothing, which is worth saying as plainly as the ones that found something. Measuring each paired passage against the middle of its own book, in the seven books whose passages pair by their own number, turns up a single outlier, and it is Gleanings 166, where the Korean edition prints its own glossary of notes. So no passage in those books has lost its text. And looking for a word written twice over finds eight, of which seven are ordinary reduplication: 높이 높이, 매일 매일, 조금씩 조금씩.

What to watch for

• No translator has reviewed any entry. Treat a high score as a strong lead, not a ruling.
• Two terms that always appear together can be mistaken for each other. The entry names the other term it shares a Korean form with, so the score can be read for what it is.
• A term the books rarely use gets a thin score. The method needs repetition and has little to say about a word appearing three times.
• Where the Korean omits a passage, nothing is invented. It is reported unpaired and dropped.
• A pairing deduced from order is weaker than one a number proves. The glossary records which is which, and can be filtered to the proven ones.