How People Translate Quickly | CoNLL-U Treebank Translation: Add Glosses Without Breaking Grammar Annotations

Translating CoNLL-U linguistic examples quickly means making an annotated sentence understandable in another language without pretending that the original grammar tree can simply be translated word by word. Universal Dependencies treebanks store token forms, lemmas, part-of-speech categories, morphological features and grammatical relationships in a precise ten-column format. A French explanatory translation of an English sentence can help learners, but replacing English word forms with French strings while retaining the old head indices and grammatical labels will usually corrupt the annotation. The safe approach separates the original analysis from sentence translations, lexical glosses and any genuinely new target-language analysis.

The primary reader job is to add clear multilingual learning support to an existing CoNLL-U treebank while protecting its machine-readable linguistic evidence. Depending on the purpose, that support may be a sentence-level translation comment, a documented gloss in the MISC column or a separately versioned annotation file keyed by sentence and token identifiers. The official format already describes optional # text_en-style sentence translations, Gloss entries, transliterations and language-specific metadata. Each has a different function. A translation is only a new treebank when the target-language tokens and syntax have been independently annotated and validated.

This guide explains the ten-column contract, why token counts and dependency structures change across languages, how to distinguish glossing from fluent translation, what to preserve in comments and alignment keys, and how to test real examples. We will examine punctuation, multiword tokens, empty nodes, morphology, code switching and the validation rules that make an annotated corpus useful for research and classrooms. The practical outcome is a reproducible translation layer that improves access without altering the evidence that the original treebank was built to represent.

First answer: choose the type of translation you are creating

A sentence translation gives a learner a fluent understanding of what the whole sentence means. A word gloss gives a rough lexical cue for a particular token or lemma. A morpheme gloss breaks a word into meaningful pieces and explains grammatical parts. Transliteration writes a word in another script, usually preserving its language. A parallel treebank, by contrast, has independently annotated sentences in multiple languages. These products overlap in educational use but cannot be treated as identical columns in a spreadsheet.

Before editing anything, state the intended deliverable. If the assignment is “help French students understand the English treebank,” a French sentence translation and a carefully reviewed glossary may be enough. If the assignment is “make a French dependency treebank,” the team needs French tokenization, lemmas, part-of-speech analysis, morphology, dependency heads and relations. Simply copying a translation into the English FORM column does not complete that work. A clear deliverable protects both time and scientific integrity.

The ten-column format is a technical contract

CoNLL-U word lines contain exactly ten tab-separated fields: ID, FORM, LEMMA, UPOS, XPOS, FEATS, HEAD, DEPREL, DEPS and MISC. FORM records the surface form; LEMMA records a base form; the part-of-speech and feature columns describe grammatical categories; HEAD and DEPREL encode a dependency relation; DEPS holds enhanced dependencies where applicable; and MISC stores additional information. Sentence-level comments begin with a hash mark, while blank lines separate sentences. The structure is designed for software as well as human reading.

A translation should not erase those distinctions. Suppose the English token “walked” is a past-tense verb. Its French equivalent might be one word or a different construction in another context, and its tense and aspect analysis may not map one-to-one onto the English features. If the translator changes FORM to a French form but retains the English lemma and morphological tags, the record becomes internally inconsistent. A parser may still accept ten columns, but that does not make the linguistic analysis true.

A small source example with stable structure

The following minimal CoNLL-U sample illustrates an English sentence with a subject, a verb and punctuation. It is written with literal tab separators in the manuscript’s code block; when exporting a sample for software testing, check that the actual file uses tabs rather than a series of spaces. The comment provides the sentence text; the dependency tree identifies the verb as root and the noun as its subject. There are no translations in the structural fields, and the sentence ends with the required blank line in the file representation.

# sent_id = demo-001
# text = Birds fly.
1	Birds	bird	NOUN	_	Number=Plur	2	nsubj	_	_
2	fly	fly	VERB	_	Tense=Pres|VerbForm=Fin	0	root	_	SpaceAfter=No
3	.	.	PUNCT	_	_	2	punct	_	_

This is an educational fixture, not a claim that all linguistic details of every English occurrence of “fly” are resolved by these tags. Its central point is structural: word line 1 depends on line 2, and line 2 is the root. A French translation such as “Les oiseaux volent” has its own token inventory and syntax. You cannot paste that sentence over the English FORM cells and keep three word IDs as if nothing changed. The target translation belongs in a language-aware support layer unless the complete French analysis is created.

Sentence-level translations can be recorded as comments

The official UD CoNLL-U documentation describes optional comments such as # text_en = ... for English sentence translation and analogous language-coded comments for other languages. This is a useful lightweight mechanism for whole-sentence understanding. It leaves the # text = ... line responsible for the original sentence to which the token analysis belongs. A French translation may use an analogous text_fr comment when that is appropriate to the project’s policy and tooling. That comment is supporting information, not a replacement for the source text.

Keep the sentence comment’s meaning clear. # text helps reconstruct or display the original utterance as analyzed, so a translator must not overwrite it with a more natural target-language sentence. If the original has unusual punctuation, a transcription hesitation or a historical spelling, the text line should remain faithful to that original under the corpus rules. A target-language translation may smooth those details for readability, but the smoothing must be visibly part of a translation, not a secret alteration of the linguistic evidence.

A word gloss is not a fluent sentence translation

The Gloss attribute described for MISC can supply an approximate translation of a token form or lemma, typically in English. It can be helpful in language-learning datasets where students need a quick lexical bridge. But a gloss is narrower than the full meaning of an utterance. The German verb ending in a specific form, for example, may contribute person and number information that a simple English dictionary headword does not express. A gloss such as “to-go” helps identify a lexical meaning, not necessarily every grammatical fact.

The official format also describes MSeg and MGloss for morpheme segmentation and morpheme-by-morpheme glossing. Those should not be confused with arbitrary multiword translation. Their delimiter conventions carry technical meaning: morphemes are aligned with segments, and grammatical tags are often abbreviated. Translating a token into an eloquent phrase and putting the phrase into a morpheme-gloss field may destroy that alignment. Define the gloss tier first, then enforce the tier’s conventions consistently.

When to use a separate translation sidecar

A sidecar file is often the safest choice for classroom translations into several languages or for experimental glosses that are not part of the official treebank. It can contain the original treebank version, sentence identifier, optional token identifier, target language, translation kind, proposed text and review status. It does not require rewriting the source CoNLL-U file. That makes a clean read-only comparison possible and prevents one school project’s glossary preferences from contaminating a public linguistic dataset.

A sidecar needs stable keys. At minimum, identify the treebank and released version along with sent_id; token-level entries also need the original token ID within that sentence. Do not assume sentence IDs are globally unique across every corpus. A token ID of 2 has meaning only in the context of its sentence, and a sentence ID reused in two releases may refer to changed annotation. Include an exact source version or content checksum so a translation can be matched to the right source without guessing by surface text alone.

An illustrative sidecar record

For the English source sentence above, a French explanation could live in a separate tabular file. A sentence-level row might say that demo-001 translates to “Les oiseaux volent.” A token-level row could indicate that English “Birds” has a French lexical equivalent involving “oiseaux,” while clarifying that the full French sentence introduces an article. The token gloss and the fluent sentence are not required to have identical word counts. Keeping both levels separately allows teachers to explain precisely why translation involves grammar rather than word substitution.

treebank,version,sent_id,token_id,language,kind,text
Demo-English,1,demo-001,,fr,sentence,Les oiseaux volent.
Demo-English,1,demo-001,1,fr,lexical_gloss,oiseaux
Demo-English,1,demo-001,2,fr,lexical_gloss,voler

This CSV is a proposed project-specific sidecar format, not an official UD standard. The version shown is an illustrative local identifier, not a claim about a real Universal Dependencies release. In a production project, document CSV quoting, Unicode normalization, licensing and identity fields. The empty token ID distinguishes sentence-level translations from word-level glosses. The sidecar’s advantage is reversibility: it can be replaced or expanded without disturbing the original analyses.

Do not change `FORM` just because another language is requested

FORM records the original word form that the treebank analyzes. When a source sentence is translated, target-language words may be reordered, added, omitted or fused. If an editor substitutes target words one by one into existing rows, syntactic categories and dependency relations cease to describe those forms reliably. This is not a minor formatting defect. A machine-learning experiment may learn false patterns from the resulting corpus, and a learner may be taught a grammar relation that is not present in the translated sentence.

The repair depends on the product. For a bilingual teaching interface, restore the original FORM and keep translations in comments or sidecars. For a new target-language treebank, tokenize the target sentence according to its linguistic conventions, annotate all relevant columns independently and validate the result. Do not treat the English tree as a reusable template for French, Arabic or Chinese. Parallel sentences can communicate similar meanings while realizing them through different grammatical structures.

Lemmas and morphological features are language-specific evidence

The LEMMA column is not a bilingual dictionary slot. It records a morphological or lexical base form relative to the annotated language. For “Birds,” the English lemma “bird” makes sense. A French translation would need a lemma appropriate to the French token, not an English lexical label placed underneath it. FEATS can contain feature-value pairs such as plural number or verb tense, but the same meaning can be expressed through different morphological systems across languages. Copying source features onto translated forms risks inventing morphology.

For example, an English noun might not encode grammatical gender in its annotation, while a French noun may do so. Conversely, a grammatical distinction marked explicitly in one language may be absent or realized differently in another. A bilingual educational note can explain the difference, but the original features should remain intact. A target-language corpus requires analysis according to the target language’s guidelines and its actual sentence. Translation memory is not a substitute for linguistic annotation.

Dependency heads are positional references, not bilingual meanings

In basic UD dependencies, HEAD refers to another word ID in the same sentence or zero for the root, while DEPREL names the grammatical relationship. If a target translation inserts a determiner or changes word order, those indices may no longer point to the intended grammatical heads. A file can retain syntactically valid integer values while assigning the wrong dependency relations. That is a particularly dangerous corruption because a quick column-count validator might not catch the semantic mistake.

The right regression question is whether every basic dependency still describes the unchanged source sentence. When the source token rows are meant to be preserved, a localization patch should leave ID, HEAD and DEPREL untouched. If new target-language analysis is being built, compute the target heads and relations from that analysis and validate them separately. Never interpret an unchanged tree diagram as evidence that a different sentence has the same structure merely because a translation is faithful in overall meaning.

Multiword tokens show why one-to-one translation fails

CoNLL-U distinguishes syntactic words from surface tokens that may contain multiple words. A multiword-token range such as 3-4 can represent a surface form that corresponds to two syntactic word lines, depending on the language and analysis. French contractions provide familiar examples. A translator who assumes that every printed space separates one annotation row may miscount tokens and damage the structure before any linguistic interpretation begins. Review tokenization conventions in the actual treebank rather than relying on visual spacing.

Consider a French sentence using “du,” which can be analyzed as components corresponding to a preposition and an article in an appropriate UD representation. Translating an English phrase into French might introduce this contracted form, but the English source tree should not suddenly gain a multiword-token line. A new French treebank would need its own tokenization and dependency decisions. A teaching translation can mention the contraction in prose, leaving the original English annotations intact. This is exactly the sort of case where speed comes from separating products rather than forcing one file to serve incompatible purposes.

Empty nodes and enhanced dependencies need their own rules

CoNLL-U can include empty nodes, whose decimal-like identifiers support analysis of elided material in enhanced dependency representations. They are not ordinary surface words to translate. Similarly, the DEPS column may encode enhanced relationships beyond the basic tree. A simplistic spreadsheet export that sorts IDs numerically or assumes all IDs are integers can reorder empty nodes incorrectly. A translator should not edit these fields when the deliverable is only a gloss or sentence translation.

If a multilingual teaching interface displays a tree with enhanced dependencies, it should use the source tree’s structure and add explanatory language around it. Show students which grammatical relation an edge represents, perhaps with an approved translation of the label in a separate UI legend. But do not replace the canonical DEPREL or DEPS values with localized words inside the file. Those are code values that parsers and validators understand, not free-text captions for a diagram.

Preserve `SpaceAfter` and original reconstruction

The MISC field can include SpaceAfter=No to indicate that the next token follows without a space, as often happens before punctuation. Changing it during localization can make the reconstructed original sentence differ from # text, and validators may detect the inconsistency. In the sample, the period is attached directly to “fly” in the original sentence, so the spacing information matters even though the French translation is stored elsewhere. A translator must not remove spacing metadata just because target-language punctuation differs.

Other source languages may use spacing conventions very different from English. Some scripts do not separate words with spaces in the same way, and typographic conventions vary for quotation marks and punctuation. These differences belong to the target translation’s presentation or a separately annotated target corpus. They do not authorize rewriting the source’s observed spacing. Keeping the original detokenization contract intact allows researchers to reproduce the sentence as it was annotated and students to compare source and target accurately.

Distinguish transliteration from translation

A transliteration changes representation between writing systems; it usually does not translate a word’s meaning. The CoNLL-U MISC conventions include Translit for surface forms and LTranslit for lemmas. These can help readers pronounce or search words written in unfamiliar scripts. A word translated into English or French, however, should not be stored as Translit because that would claim a relationship of script representation rather than lexical meaning. The distinction matters for applications that use transliteration to align characters or teach pronunciation.

For a lesson involving an Arabic word, a Latin-script rendering and an English meaning are both useful, but they answer different questions. A student may need to know how the word is pronounced, what it means in context and which grammatical features it expresses. Store or show those as separate layers, with explicit conventions. Do not copy the translated English meaning into the transliteration field simply because it uses Latin characters. Reversible terminology begins by giving each field a precise job.

Code-switching is not a translation request

A sentence can contain words from more than one language. Universal Dependencies provides conventions for marking foreign material and, in certain cases, token-level language using the Lang attribute in MISC. That is an annotation of the original mixed-language utterance, not a request to translate its foreign token into the surrounding language. Converting the original code-switched form into a monolingual paraphrase may remove the very phenomenon the treebank was designed to study.

A teaching translation can explain the whole utterance in a learner’s language while retaining the code-switching evidence. The reviewer should understand whether an embedded term is a proper name, a borrowed lexical item, a quoted foreign expression or a switch requiring language-specific analysis. A general translation engine may normalize the whole sentence and erase the switch. Preserve the annotated source first, then describe the code-switching effect in the target learning material instead of changing what the speaker actually said.

Translate grammar labels for readers outside the file

A class learning dependency grammar may benefit from French descriptions of nsubj, obj, root, case and other relation names. Those descriptions belong in a user-interface glossary or teaching notes, not as replacements for the standard relation codes in DEPREL. The code nsubj has a defined grammatical role in the annotation scheme, while its natural-language explanation can vary with the learner’s background. Keeping them separate lets an English and French classroom use the same dataset and automated queries.

A useful glossary entry contains the canonical code, a plain-language definition, a positive example, a confusing near-neighbor and the approved target-language explanation. For nsubj, do not simply translate it as “noun before the verb,” because subject position and category vary across constructions and languages. The translation should describe a grammatical relation rather than a superficial word-order shortcut. This is where localization becomes educational: students learn to distinguish labels as analytical concepts from the surface words in one sentence.

Align sentences first, not guessed word positions

When attaching translations across files, link by stable sentence identity and exact corpus version. Sentence order alone is fragile. Editors may shuffle examples, split a file or update annotation, shifting every later row. A sentence’s string text is also an unreliable key because identical utterances can occur in different contexts or documents. Use the official sent_id where available together with treebank name and release, then create a documented mapping to translations. Detect duplicate or missing identifiers before making a final export.

Word-level alignment requires even more care. A source word can map to several target words, no target word or a phrase that carries a meaning distributed across the sentence. A sidecar should express that ambiguity rather than force every token into exactly one target slot. If the intended learning feature needs aligned spans, define span identifiers or alignment sets and review them independently. A spreadsheet that assumes source token 4 always corresponds to target token 4 can be convenient to create and disastrously misleading to use.

Machine translation can draft, but not validate syntax trees

Automatic translation can provide a first draft of a fluent sentence translation, especially for familiar modern language pairs. But the model may add subjects, normalize discourse particles, invent punctuation or silently resolve ambiguity. Such changes may be acceptable in explanatory prose yet unsuitable if the translation is meant to preserve the source’s formal linguistic features. Give the translation engine the purpose: beginner comprehension, literal comparison, morphological gloss or publication-ready parallel sentence. The output constraints differ sharply among those tasks.

Never use translated sentence fluency as evidence that its dependency tree is correct. Syntax annotation needs language-specific linguistic judgment and validation. An English parser and a French parser may agree on a broad clause meaning while assigning different token spans or relations. If the project aims to compare grammar across languages, those differences are the research result, not errors to smooth away. A well-designed workflow allows translators to draft comprehension layers quickly while trained annotators handle structural analysis separately.

Review examples with an annotation-aware screen

Show the reviewer the original # text, the token table, the source dependency graph where useful, the proposed target-language sentence translation and any lexical glosses in separate panels. A reviewer should be able to identify the exact source token associated with a gloss without editing the structural columns. Include the lemma and part-of-speech information so the reviewer can distinguish a noun sense from a homographic verb sense. Show enough surrounding context to resolve pronouns and ellipsis, but avoid exposing unnecessary personal details from a corpus.

The interface should make uncertain alignments visible. If an idiom translates as a phrase with no direct word-by-word correspondence, mark the gloss as explanatory rather than pretending every component maps cleanly. For teaching, an explicit note such as “This English idiom is expressed by a different construction in French” is often more useful than five false dictionary equivalents. Human review focuses on preserving the phenomena the source demonstrates, not on making all source and target word counts match.

Data licensing, attribution and privacy remain relevant

Treebanks often have specific licensing and attribution requirements, and some source sentences may have been collected under conditions governing redistribution. Translation does not remove those obligations. Before packaging new language annotations, inspect the treebank’s license and its original source notes. Determine whether translations can be distributed, whether the original text must accompany them, and how contributors should be credited. A permissive format specification does not automatically grant permission to reuse every corpus encoded in that format.

Privacy can be a further issue if the material comes from conversations or documents involving identifiable people. Avoid posting real learner data into an external translation service without authorization. Synthetic test sentences are adequate for verifying tabs, columns, comments and alignment logic. If a project must process restricted text, use its approved environment, record who can access the translations and avoid leaking content through logs or public review links. A corpus can be technically valid and still be unsuitable for unrestricted publication.

The first QA gate: plain-text structural integrity

CoNLL-U files should meet the official rules for encoding, line endings, sentence separators and tab-delimited fields. After a translation pass, check that every noncomment token line still has ten fields and that sentence boundaries remain intact. Compare all source token lines byte-for-byte when the project is only adding comments or external translation annotations. A small formatting change may be harmless at the surface level but significant to downstream tools that expect consistent spacing or token IDs.

When a new comment is inserted, retain the relationship between that comment and its sentence. Put it with the correct sentence’s metadata, not at the end of the file where it can be misread as belonging to another example. Be mindful of the official guidance to preserve comment order unless a tool intentionally modifies it. If a program rewrites the entire file, test that it has not changed normalization, reordered enhanced dependency fields or removed custom metadata while inserting a single translation.

The second QA gate: official UD validation

Universal Dependencies maintains validation rules beyond simple column counting. They test structural format, grammatical inventories, dependency trees and other language- and release-specific requirements. Use the official validation guidance and, when appropriate, the project’s current validator rather than relying on a generic CSV parser. A file may be ten-column correct while containing an illegal relation, a broken dependency tree or a mismatch between the displayed source sentence and token spacing. Translation workflows should not weaken those protections.

It is important to distinguish local tool success from official release acceptance. The UD project notes that local validator versions or configurations can differ from the current release process, and that repository-level checks matter too. If a team intends to submit a modified treebank to the UD ecosystem, follow its current contribution and validation rules. If the team only publishes a sidecar of translations for classroom use, validate the sidecar schema separately and verify that it joins correctly to the untouched official source release.

The third QA gate: semantic equivalence of teaching content

Structural validation cannot tell whether a French translation accurately represents an English sentence. Build a parallel quality rubric: speaker and participant identities, action, negation, quantity, tense/aspect where relevant, register, idioms and omitted context. For a sentence with “did not,” omission of the negative particle is a major factual error even when the target sentence is elegant. For a sentence containing a modal such as “might,” an overly certain target statement erases linguistic information learners may be asked to analyze.

Glosses need a different rubric from sentence translations. They should identify the intended lexical or morphological contribution in context, and they may be intentionally less fluent. Ask whether the gloss matches the lemma’s sense and the token’s grammatical role. Do not demand that a one-word gloss reproduce every nuance of a whole sentence. Reviewers should explain which tier they are rating; otherwise a useful literal gloss may be rejected for sounding awkward, while a smooth paraphrase may be accepted despite hiding a structural contrast.

Diagnose a broken tree after replacing word forms

If visualization suddenly shows edges attached to unexpected words, inspect whether translated target words were written into the original FORM rows. Count source and target tokens and compare the ID, HEAD and DEPREL columns with the frozen original. If the source analysis is intended to remain, restore all structural rows from the verified baseline and move the translation into an approved annotation layer. The repair is not to adjust a few head numbers until the picture looks plausible; the entire target-language analysis would need linguistic review if that is the intended new deliverable.

This failure often originates in a spreadsheet whose columns are labelled only “word,” “meaning” and “relation.” Add exact CoNLL-U field names to the export schema and lock structural columns for translators. Provide a target translation column that cannot accidentally serialize as FORM. Version-control the exporting code and exercise it on a sentence with a different token count after translation. If the pipeline cannot represent a valid two-to-three word mapping, that is a reason to redesign the pipeline, not a reason to force the language to fit.

Diagnose a word gloss placed on the wrong token

Suppose the French gloss for “fly” appears beside “Birds.” The underlying CoNLL-U may be completely unchanged; the defect is in the sidecar’s sentence-token join. Verify the treebank version, sentence ID, token ID and treatment of punctuation or multiword-token ranges. If a preprocessing step silently dropped punctuation, positional alignment may have shifted. Rebuild the join using explicit identifiers rather than line numbers or array positions, and write a test asserting that every gloss links to the intended source form and lemma.

A useful review screen displays both the key and source text. This lets a human catch a record that technically resolves but points to an unexpected token because an older corpus version reused the ID. Store the source token’s form or a content hash as an additional assertion, not as the sole identifier. The goal is to make a mismatch visible early rather than letting it contaminate a batch of thousands of examples.

Diagnose corruption introduced by sorting

A program that sorts token identifiers as strings may put token 10 before token 2, or mishandle empty-node identifiers such as 5.10 and 5.9. That can make a CoNLL-U file appear scrambled after a harmless-looking CSV round trip. Use a parser aware of the format’s ID types and ordering rules. Better still, leave the source token lines untouched during translation and keep the sidecar in its own order. A file’s line order reflects the analyzed sentence sequence and is not an arbitrary table layout.

Similarly, multiword-token lines with range IDs have a particular relationship to their member word lines. Sorting them independently can separate the surface form from its analysis. The repair is not an ad hoc alphabetic sort by FORM; it is restoration from the source and a format-aware transformation. Include fixtures with more than nine tokens, at least one multiword token and an empty node if your project supports enhanced graphs. A test built only from three-word sentences will miss these failures.

Diagnose a translated `# text` that no longer matches the rows

If a validator flags a source-text reconstruction mismatch after localization, inspect the sentence comment. A translator may have replaced # text = Birds fly. with a French sentence while leaving the English token rows intact. The comment now contradicts the analyzed surface words. Restore the original # text, add the French explanation as a language-specific comment or sidecar, and rerun validation. The localized sentence is useful, but its field placement was wrong.

The same failure can occur when punctuation is “corrected” according to a target language while the original data lines are preserved. Source punctuation is evidence, not necessarily an editorial mistake. In an educational interface, show the translation’s punctuation in a distinct display region and retain the original SpaceAfter metadata. This supports comparison: learners can see where languages differ in word order and typography without mistaking target conventions for source observations.

Diagnose confusing translation of an idiom

A sentence like “She kicked the bucket” can be literal or idiomatic depending on context. A dictionary gloss for each word may produce the literal interpretation even when the full sentence means something else. Conversely, a fluent idiomatic translation may not preserve the lexical material relevant to a syntax lesson. Decide what the teaching exercise asks students to learn. Provide a contextual sentence translation and, where useful, separate lexical glosses plus an explanatory note about the idiom’s meaning and construction.

Do not edit grammatical dependencies to make an idiomatic interpretation align with a translated phrase. The tree describes the source utterance as annotated. If researchers disagree on a construction, document that as an annotation issue to be handled by the corpus owner. The translator’s task is to help readers understand both the form and the meaning faithfully, not to settle linguistic theory through a hidden reannotation.

Diagnose target-language font or segmentation problems

The CoNLL-U file can be structurally correct while a teaching website renders a target script poorly. Missing glyphs, broken shaping, mixed text direction or incorrect line wrapping may make the gloss unreadable. Test the display with the target language’s actual script, punctuation and combining marks. A string stored in UTF-8 is not guaranteed to display correctly if the application chooses a font lacking the needed characters. Keep visual rendering checks separate from CoNLL-U format validation.

For languages with right-to-left scripts, do not assume that punctuation and token identifiers will follow the same visual order as an English example. The annotation table’s numerical order is a technical structure; the rendering layer must handle bidirectional text safely without rewriting token IDs. A successful multilingual corpus interface preserves both the source analytical order and the target’s natural reading direction. When a learner copies text from the interface, verify that the logical character sequence remains usable.

A transfer exercise: Japanese examples with English translations

A class has a CoNLL-U file representing Japanese sentences with segmentation and dependency annotations. Teachers want to give English-speaking students whole-sentence translations and approximate word glosses. An editor proposes to replace Japanese FORM strings with English equivalents, keep the original HEAD values, and rename UPOS labels to English classroom words. Another editor proposes to preserve the Japanese file and create sentence-linked English translations and carefully documented glosses. Which approach is sound, and how would the class verify it?

The second approach is sound. Preserve Japanese source word forms, lemmas, UPOS codes, features and dependencies. Add approved English # text_en comments where appropriate or maintain an aligned, versioned sidecar; use gloss conventions carefully, distinguishing lexical meaning from morpheme analysis. Run the format and relevant UD checks on any altered source file, then check that translation records join to the correct sentence and token IDs. Review meaning on idioms, negation and ellipsis. If the project later needs an English dependency treebank, annotate actual English sentences separately rather than borrowing the Japanese structure.

Build a practical review checklist

At the start, record the exact treebank name, version, source license, permitted use and intended translation tier. Freeze the source annotations and define a stable key strategy. Ensure that no translation is written into FORM, LEMMA, FEATS, HEAD, DEPREL or DEPS unless the project is explicitly creating and reviewing a new annotated corpus. Distinguish # text, target-language sentence comments, glosses, transliteration and morpheme segmentation. Confirm that any sidecar’s language labels, delimiter rules and review status are documented.

Before release, parse representative files, check tab counts and sentence boundaries, compare source rows with the baseline, run suitable UD validation, test alignment keys and examine the learning interface with actual target scripts. Review boundary sentences containing negation, multiword tokens, ambiguous lemmas, code switching and longer-than-nine token sequences. Preserve reviewer provenance and document known imperfect glosses. The result should be reversible: a user can remove the translated layer and still recover the exact source treebank, with its original grammatical claims intact.

Limits: translation cannot replace linguistic annotation

Even excellent sentence translations do not create a parallel treebank automatically. Two languages may express the same proposition with different word counts, morphology and syntactic organization, and some meanings cannot be aligned token by token without explanatory judgment. Automated alignment may help researchers propose mappings, but the output is a hypothesis that needs task-specific evaluation. A fluent translation may be ideal for comprehension yet unsuitable for fine-grained analysis of negation, scope, agreement or information structure.

Likewise, the CoNLL-U format does not prescribe a universal multilingual teaching interface. It specifies an annotation representation and documented conventions for supplementary information. A sidecar designed for a classroom can be useful without being an official extension to the UD standard, provided it is labelled honestly. If a research team plans to contribute new data to an official release, it must follow the project’s current validation, annotation, licensing and contribution rules. Do not confuse a successful local demonstration with formal inclusion in a treebank release.

Summary: keep the annotated sentence and its translation in view

The fastest reliable way to translate CoNLL-U examples is to identify which layer a learner needs, then add that layer without changing the original grammatical evidence. Whole-sentence translations communicate meaning; lexical and morpheme glosses reveal local structure; transliteration supports script access; and a true target-language treebank requires its own annotation. Stable sentence and token keys, format-aware tools and versioned review allow these products to work together without being mistaken for one another.

This approach teaches a valuable idea about language itself. A translation can preserve a thought while changing the number, order and grammatical form of the words expressing it. A dependency tree describes the actual words of one analyzed sentence, not an abstract sentence that can be moved unchanged between languages. Keep those two truths visible, and multilingual corpus work becomes both more accurate and more educational: students gain access to meaning without losing the linguistic evidence that makes the data worth studying.

Frequently asked questions

Can I replace `FORM` with the target-language word and leave the tree alone?

Generally no. The tree and morphological annotation describe the original tokenization and sentence. Translation usually changes those relationships.

What does `# text_en` mean in CoNLL-U?

The official format describes it as an optional English sentence-level translation, analogous to language-coded comments for other languages. It does not replace the original # text.

What is the difference between `Gloss` and `Translit`?

Gloss provides approximate meaning; Translit represents a form in another writing system. They should not be conflated.

May a gloss contain more than one word?

The official Gloss convention discusses joining multiword equivalents with hyphens. Review the project’s exact MISC policy before writing them.

Can I change `DEPREL` to a translated grammatical label?

Not in a standard CoNLL-U source file. Keep the canonical relation code and translate its classroom explanation in a separate interface glossary.

Why do French translations sometimes have more words than English originals?

Grammar, articles, pronouns, contractions and idioms can produce different tokenizations. A one-to-one token mapping is not guaranteed.

How do I handle multiword-token ranges or empty nodes?

Use a format-aware parser and preserve their source structural meaning. Do not treat all IDs as ordinary integers or sort them lexically.

Is a fluent machine translation sufficient for a parallel treebank?

No. A target-language treebank needs independently established tokenization and linguistic annotations, with appropriate validation.

Is an external sidecar an official UD standard?

No. It is a local integration strategy that can be safe and useful if its schema, version keys and limitations are documented.

What is the minimum useful QA comparison?

Confirm the original CoNLL-U token lines and annotation remain unchanged, verify every translation’s source key, and check semantic accuracy plus rendering in the target language.