Reading a book nobody has glossed

Between pasting a text in and finishing its glosses lie weeks, and for those weeks every unglossed chunk is a phrase you cannot get past. Three things help, all on this machine, and none of them is a gloss:

a dictionary

gives you the words: what each one can mean, and how the word on the page was reached from the headword it is a form of;

a translation model

reads the sentence your chunk sits in, and marks the part of its reading that seems to be your chunk’s;

sentences somebody has already translated

— a real sentence, written by one person and translated by another, sharing the rare words of your chunk — settle which of the senses on the list is the live one.

They are installed once, with buttons, on the reading help page (Settings, /settings/reading-help/), which Get a dictionary for this language →, in the ⚙ panel’s Looking a word up, opens; the Lookup and languages section of this guide goes through it. A reader with none of them is exactly the reader it always was.

1. The dictionary switch#

Once anything is installed for the book’s language, the header shows dictionary. It is off until you turn it on — then it is remembered, for this book, in this browser — and its tooltip says what it will do: look the words up, show a sentence somebody translated, translate the line — where nothing is glossed, or as much of that as is installed. With it off, nothing is fetched, nothing is translated, and the page costs exactly what it did before.

With it on, click any chunk that has no vocabulary line — in any mode, hover or not — and its cloud opens with the panel in it. (In hover mode, pointing is enough.) A chunk somebody has glossed is left alone: the panel exists for the chunk nobody has got to yet, and a finished edition looks exactly as it did.

English is explained in English. Every dictionary here is the English Wiktionary’s, which explains the words of every language in English — for Persian a translation, but for English itself a definition, in the very language you are learning. So a book in English has two more switches, in the ⚙ panel’s Looking a word up: Show the dictionary’s definitions (off until turned on) shows Wiktionary’s definitions under each word, with their labels (transitive, countable, slang), the first three at once and the rest behind N more definitions; and, where a translation model into the language of your glosses is installed, Translate the definitions into Italian (named after that language) puts each definition into it, underneath — a machine’s reading of a definition, labelled so. Off, an entry still gives the headword, how it is said, its part of speech, how it was reached and a verb’s forms.

2. The panel#

The panel is tinted and ruled off from the gloss, and holds up to three blocks, each headed with what it is and ending with where it came from:

dictionary — not a gloss. Each word of the chunk, and under it every entry the dictionary has: the headword and its romanisation, the part of speech, and the senses — the likely ones first, obsolete and rare ones last. When the word on the page is not the headword, a line says how it was reached (found without the marks and the case, for instance), which is the line that teaches. A verb carries a line of its principal parts. A word it cannot find says what it tried: not found (tried …). At the foot, the source and its licence.

a machine’s reading of the sentence — not a gloss. The whole sentence the chunk sits in, translated by the model on this machine. A chunk is a fragment by construction, and a fragment translated alone comes back as one, so it is the sentence that is read; the words of the reading that seem to be your chunk’s are marked in it — in colour and bold where the dictionary accounts for at least half of the chunk’s words, with a dotted underline where it accounts for fewer — and a line under it says what the mark stands on, this chunk is likely: “…” — مثل → parable · ایران → Iran, or part of this chunk is likely, or that nothing matched and nothing is marked. The mark is a guess from meaning, never from position: word order is exactly what a translation changes. Where the chunk is the whole sentence, there is nothing to mark, and the heading is just a machine’s reading — not a gloss. The model and its engine are named at the foot. There is no button to use it as a gloss, and there will not be one.

a sentence somebody translated — not this one. Sentences from a translated corpus (Tatoeba) that share the rare words of your chunk: both sides, the shared words marked on both, and shares … naming them. Load more fetches five more, for as long as there are more. It is not your line and never pretends to be, which is exactly its use.

A panel with nothing to say says so: the dictionary has nothing for these words.

It gets ahead of you. While you are not asking anything, the reader looks up the unglossed chunks of the next ten subparagraphs, and reads their sentences with the model, so that the panel opens already filled. It waits for a moment’s quiet after your last scroll or keypress, and stands down the moment you open a cloud yourself — and it does nothing at all while the switch is off.

3. The sources sidebar: taking it up into a gloss#

The panel is for reading; nothing in it writes. Where the gloss is actually written — the chunk sheet (Writing a chunk) — the same sources come with you: ⊕ sources in the sheet’s head opens a column on the left of the fields (the picture on that page shows it). The fields do not move; the sheet grows to the left. On a narrow screen the column goes under the fields instead, and the sheet scrolls down to it. The column stays open until you close it, remembered for this book.

It is offered on every chunk that has gloss slots, written or not — correcting a line is as much of the work as writing one — and whether or not the dictionary switch is on. Its four blocks, in the panel’s order:

Every button appends — a vocabulary line is built up entry by entry — and leaves the cursor at the end of the field. It writes into a box, never into the book: nothing is saved until you press save chunk, down the same route, with the same checks, as anything you type.

A verb goes in as \vb. Where the dictionary’s entry is a verb the language knows how to conjugate, the row shows the line of its forms, and the button is \vb → vocabulary: the verb with its principal parts, in this language’s order. Where the book has already glossed the same verb, the row first offers \vb as this book glosses it →, so the twenty-first occurrence matches the twenty before it; the dictionary’s own is then the dictionary’s \vb → vocabulary, and is not drawn at all when the two are the same. A \vb the dictionary could not complete still goes in, with its gaps listed under the row, one to a line, headed to fill in: — and its button has a dashed border. Where the verb is another verb’s than the headword (German’s stand … auf going in as aufstehen, a reflexive going in as sich freuen), the row says so with an arrow.

A verb of two words. Persian’s compound verbs (labxand zadan, to smile), Turkish’s, Hindi’s conjunct verbs and French’s verbal locutions are one vocabulary entry, not two. For them the row has a second button, compound verb → vocabulary (conjunct verb in Hindi, verbal locution in French), that puts the pair in whole; its tooltip says what it will write.

What is always left to you: the meaning is the dictionary’s first sense, not necessarily the one your sentence wants; where two auxiliaries are possible, both are printed — strike the one the sentence does not use; every hit that could be a verb gets a \vb, the unlikely ones included. A block with nothing to offer says so in its own words (the dictionary has nothing for these words, no translation model for this pair).

4. What the verb draft does in each language#

A verb’s draft is built from the dictionary’s conjugation table, so it is as good as that table, and the tables of the eleven languages have their own gaps. It is always a draft for you to correct, and where it could not fill a slot the button says which. In a video, where the phrase holds a form that is none of the three a \vb prints, the draft also names that form after the entry, outside it — ; here \pw{vint} \textit{vẽ} — and what to call it (past historic, say) is yours to type. What each language’s draft proposes, and what is left to you:

Persian.

The infinitive and the two stems, each with its sound. Where the book already has a \vb for the same infinitive, the book’s own is offered first. The stems are Wiktionary’s, from the literary Iranian table where there is one, and need not be the book’s — it has dah for دادن where the series writes deh — so check them against the conventions, and trim the meaning to the one sense the text uses. For a compound the dictionary knows (فکر کردن and عوض کردن — not every one it should) the draft is the light verb’s \vb with its meaning already empty, and the compound verb button puts the whole pair in, saying in so many words that لبخند زدن is one verb written in two words and not two entries: \vb{زدن}{zadan}{زن}{zan}{زد}{zad}{}\bw{لبخند}{labxand}{to smile}, the \bw run straight onto the \vb with no ; between them, which is what makes the pair one entry in the line (in a video, with the colloquial present in the light verb’s parenthesis where there is one). The plain \vb button stays beside it — the light verb alone, still drawn unfinished — and the row goes on naming the \bw under to fill in, so you see what the entry is about to become before you press. Where the dictionary gave no meaning the compound still goes in, with that slot empty and the button dashed; where it does not know the compound at all (گمان کردن or معلوم شدن) the light verb comes with its own meaning, and emptying it is yours. A preverb is hyphenated only where the dictionary marks one (برگشتن, برداشتن, درآوردن, فراگرفتن and a few more): برخاستن comes as barxāstan, and the hyphen is yours to put in. A preverb verb the text writes apart (بر می‌گردم, در آورده) is offered whole, as برگشتن and درآوردن. In a video the draft adds the colloquial present wherever the dictionary’s Tehrani table spells it differently — 18 verbs, among them mi-gam, mi-ram, mi-šam and mi-dunam — and you take it out where the speaker does not say it.

Arabic.

The vowelled perfect with its form, the imperfect, a masdar and a governed preposition where the dictionary has one, each transliteration respelt in the book’s scheme (the dictionary writes ʔ ʕ ḵ ḡ and an initial hamza). Where the dictionary lists several masdars — it does for about one verb in six — the draft has the first, which is often the one for another sense (صِلَة where to arrive wants وُصُول); the meaning and the preposition are those of the sense the dictionary gives first, and its meaning runs to a line of synonyms; and a hit reached through an unvowelled word may be the wrong verb altogether. A masdar the dictionary does not know, which it does not tell apart from a verb that has none, goes in blank, and the button says so. In a video the draft comes without the vowel marks, as a video’s line is written; the transliterations keep the vowels.

Italian.

The entry for a verb it recognises, the auxiliary and an irregular passato remoto included, and — for a form the dictionary only knows as a compound (alzandosi, dimmelo, farlo) — the entry of the verb it is a form of. The meaning is the dictionary’s first sense cut to one equivalent, which is often not the one the chunk needs (dovere comes as to owe). Where a reflexive pronoun stands before the word and the dictionary lists the two together as a form of the pronominal verb (si alzò, mi alzo, me ne vado, si è alzato), it proposes the pronominal verb (alzarsi, andarsene) — but with si it cannot know a reflexive from an impersonal or passive one (si dice, si vendono case are dire, vendere), and the button says so. The auxiliary is the one the dictionary’s head line gives, and that is sometimes wider than usage: sapere and usare come with avere/essere, the essere for a sense the text will hardly have (to taste of, to be in fashion) — trim it to what the chunk uses.

Japanese.

The three forms with the rōmaji the dictionary gives them (the -te form is its past with た made て: 書いた kaita → 書いて kaite), the class it records, checked against those forms — so 帰る, which looks ichidan, comes out godan, as its stem 帰り says — tr. or intr. only where it records one (about half the verbs; the rest are left without, not guessed), and hon. or hum. where the sense it prints is tagged so. A verb the chunk writes in another of the dictionary’s spellings is proposed in the chunk’s: 飲んだ gives 飲む, though the dictionary files it under のむ. The meaning is the dictionary’s first sense, which is often a definition rather than the one equivalent wanted, or not the sense of the passage; an auxiliary after the -te form (いる in 書いている, しまう, みる) is proposed as a verb of its own, and is glossed instead as the construction, \textit{-te iru}.

French.

The three forms with their sounds, the auxiliary, an irregular nous form and an irregular future, all read from Wiktionary’s conjugation table. The sounds are Wiktionary’s pronunciations written into the book’s scheme by rule: they follow the scheme, and say what Wiktionary says — where it gives aurai as /ɔ.ʁe/ the draft has òré. The meaning is a dictionary’s first sense rather than the text’s. A verb is taken as pronominal only where the chunk shows the pronoun — se, s', or me, te, nous, vous agreeing with the subject (je me lève, not il me regarde) — and then the draft is the pronominal verb, aux. être and a reflexive sense; il s'en va is s’en aller. Where Wiktionary lists two forms (paye and paie, assois and assieds) the draft offers the first. A verbal locution the chunk makes — j'ai peur, il fait attention, even with the negation (je n'ai pas peur) — is recognised as the locution it is: the verb comes with its seventh argument already empty, the row names the \bw still to add (bw for peur after it: to be afraid), and a button of its own, headed verbal locution, puts the pair in as the one entry the file asks for, \vb{avoir}{avwar}{ai}{é}{eu}{ü}{}\bw{peur}{peur}{to be afraid}, with the \bw run straight onto the \vb. The noun’s sound is the noun’s own pronunciation respelt; where it is missing the slot stays empty. It is found only where the noun stands right after the verb in the same chunk and the dictionary has the pair as a verb page of its own; a locution it has not got comes as the plain verb with its own meaning, and emptying that is yours.

German.

The three principal parts, the third person where it changes, the auxiliary, and the case or preposition where the dictionary gives one for the meaning; sich in front where the chunk holds the reflexive pronoun (or the verb has no other use); and a separable verb put back together — at stand, the entry of aufstehen — where its prefix ends the clause in the text that was looked up. It offers no \vb for hat, ist or wird where the clause ends in the participle or the infinitive they make a tense with: that pointer is yours to write. A verb the dictionary gives both auxiliaries says aux. haben/sein until you strike one for the sense in the text (fahren, and liegen, sitzen, stehen, whose perfect the south makes with sein); a preposition that takes either case is printed auf + … where the dictionary does not say which; a prefix outside the text that was looked up is not seen, and the finite verb then comes as its plain stem (stand, stehen); the sound slots are empty, for you to fill where a pronunciation line is given; and the meaning is the dictionary’s first sense rather than the text’s.

Turkish.

The aorist from the dictionary’s head line, the present from its conjugation table, the first equivalent of its first sense, and the government where the dictionary tags that sense — which is seldom, 78 verbs in 2 559 (bakmak -e, korkmak -den, evlenmek ile), so most of it is yours to add, and beklemek’s (-i) always is. It stops there: the segmented form the text has (; \textit{…}) is yours as well. A homograph comes as two entries, each with its own forms — yenmek to defeat, yener, and yenmek to be eaten, yenir — and you keep the one the text means. For a compound it proposes the light verb’s \vb with the empty meaning, and the row names the \bw still to add, with the compound’s own meaning and government (bw for teşekkür after it: to thank (-e)); the compound verb button puts the pair in as the one entry it is — \vb{etmek}{}{ediyor}{}{eder}{}{}\bw{teşekkür}{}{to thank (-e)} — and what it writes in the \bw is the compound’s meaning, the one the dictionary’s entry is for: an edition that would rather gloss the noun itself (thanks (-e)) trims it there, as it trims every other draft. Nothing is romanised: both sound slots stay empty, and the segmentation is yours. The compound is found only where the noun stands right before the verb in the same chunk, which is where the chunking rules put it. A derived verb the dictionary has only as a note on its base (yapılmak, passive of yapmak) comes as its infinitive alone, the forms and the meaning left to you — never with yapar under it; and about one verb in twelve has no conjugation table in the dictionary (ilerlemek, incelemek, hedeflemek) and comes without its present, or without both forms. Where the dictionary contradicts itself (hafifletmek: its head says hafifletir, its table hafiflediyor) the button says why.

English.

The three principal parts with their sounds, and the present of be, have, do and say, the four whose present nobody could build from the plain form. It never offers a \vb for a modal, nor for the particle of a phrasal verb (the down of put it down is not the verb to down). Where the dictionary lists two pasts it takes the one the chunk has, and otherwise the first listed, which is the American one (dreamed, learned). A phrasal verb is proposed whole when its particle ends the chunk (came in, put it down, what he was waiting for); one with words after its particle (gave up smoking) is proposed as the plain verb, for you to lengthen. The sound of a regular past is worked out by the book’s rule for endings; an irregular one is the dictionary’s own transcription of that form, and a phrasal verb with none of its own is said as its verb and its particle (give up gɪv ʌp). The dictionary’s transcription of many a word is British only (stop /stɒp/, wrote /ɹəʊt/, and was and were have no American one at all), and such a slot is left empty rather than converted from one accent into the other; and the meaning is the dictionary’s first sense, which is not always the one in the text, nor built of words commoner than the headword.

Hindi.

The stem, the perfective and the ने mark, each with its transliteration. The forms are the dictionary’s conjugation table and the sounds are its romanisation of each form, respelt into the book’s scheme (its tilde is ṁ, its ŕ is ri, its ṣ is ś), so the schwa is right form by form — samajhnā but samjhā. The ने mark is a proposal and never a fact, and the meaning is the dictionary’s first sense rather than the text’s (रखना comes as to keep where the text may mean to put). The chunk decides two things. In a conjunct the light verb comes with no meaning and the row names the \bw still to write; the second verb of a compound, the रहा of the progressive and the था or है after a participle get no \vb at all, only the \dw they always had, because what they add is yours to say.

Spanish.

The two forms, and an irregular participle or future, read off the dictionary’s own conjugation rows and never made up. It reads the chunk for the pronominal verb: me voy, te quedas, nos casamos, siéntate and irse give irse, quedarse, casarse, sentarse. se alone gives the -se verb too, but the button asks whether it is — se fue is irse, while se dice and se venden casas are the plain verb with a passive or impersonal se, and nothing in a dictionary tells them apart. A clitic in front of an auxiliary or a modal (se ha ido, me tengo que ir) and an infinitive with me or te on it (irme looks like llamarme, which is to call me) get the plain verb: those are yours. haber comes as hay where no participle follows it (había una vez, ha habido problemas) and as the auxiliary where one does (había llegado). A fixed phrase is never reached: the dictionary is asked one word at a time, so se dio cuenta comes back as dar, and darse cuenta is yours to write. The meaning is the dictionary’s first sense and not the text’s; for a pronominal verb, the first sense marked reflexive, which is often not the one meant (ponerse comes out to turn on). ser and estar come with the division between them in its own words, without the two italics the conventions put round the words that carry it, which are yours to add. Where the dictionary lacks a slot (a defective verb has no yo: manir), or cannot say whether the participle or the future is irregular, the button says which.

Chinese.

A \vb for a verb the dictionary marks as separable or as a verb with a complement, and a \dw for everything else. The characters are the dictionary’s own simplified spelling, the pinyin its standard Mandarin reading, the split always the plain A了B. The dictionary marks fewer verbs than there are: the commonest it misses — 散步, 跑步, 唱歌, 请客, 听懂, 看懂, 回来, 进来, 学会, 记住 and some fifty more — are kept by hand in lib/lang/zh.verbs.json, which is also where its one common mistake is put right: it marks 想到 as verb-object, and 想了到 is not Chinese (想不到 is). A verb in neither — 吃完 and 找到 are not in the dictionary as words at all, and come back as 吃 and 完 — is offered as \dws, and the \vb is yours to write. A word the dictionary has only in another lect — Cantonese 揸车, whose perfective is 揸咗车 — gets no \vb, because the 了 of the split is Mandarin grammar; nor does a word with no standard pinyin, which is how the dictionary says a word is not standard Mandarin, nor a split that belongs only to an old reading of a word whose plain reading does not split: 知道 “to know the Way” is marked separable, 知道 “to know” is not. A verb that has come apart in the sentence (结了婚, 见过面, 睡了一觉) is not found as one word, and a potential the dictionary lists as a word of its own (看不见, 听不懂) comes back as that word, with a \dw: in both cases the \vb is the dictionary form’s, and yours to write. Where the dictionary’s only sense for a verb is a pointer (聊天儿 is “erhua form of 聊天”) the meaning is left empty and the button says so.

5. Ask LLM#

The last block hands the sentence to an external chatbot — any one, in another tab — and brings its answer back:

1. Ask LLM copies a prompt to the clipboard: the sentence, the three sentences before and after it, the dictionary’s results for the chunk, and every translated sentence the corpus has for it (it gathers them all first: Preparing Ask LLM with all Tatoeba examples…). The machine’s own reading is deliberately never in it. Prompt copied. Paste it into a chatbot, then paste its answer below.
2. Paste the chatbot’s translation into the box, and press Use translation (or Ctrl+↵). Translation ready for this whole sentence.
3. It is drawn like the machine’s reading, with this chunk’s share marked and the same buttons: the marked words → meaning, the whole sentence → meaning.

The answer is kept for the sentence, so the other chunks of the same sentence reuse it (Reusing the translation pasted for this sentence.) and only the mark moves. If the clipboard cannot be reached, the block says so and Ask LLM tries again.

Ask LLM helps you write one chunk’s gloss yourself. To have an LLM write the glosses of a whole stretch — every chunk nobody has glossed, from one sentence to a chapter — use the header’s gloss with an LLM instead (Glossing a stretch with an LLM): its answer goes into the book, where nothing here does, but only into chunks that have no gloss, and only whole.

6. The gloss and the fluent translation#

What the machine’s reading and Ask LLM bring is a fluent translation of the whole sentence, and that is what they are for: to understand it. The meaning you save under a chunk is another thing, a gloss: it says what that chunk’s own words say, in the order of the text, so that read in a row the meanings are stiff, and may be poor English, and you can point from each word of them to the word it renders and see how the language builds its sentence. That is why the marked words → meaning and the whole sentence → meaning only start the box: they put in a share of a fluent sentence, and it is yours to cut down to what this chunk’s words say, and to bring into line with the vocabulary line. It is also why Ask LLM asks the chatbot for a translation, and stays so, while gloss with an LLM asks it for glosses (What the meaning says).