Frequently asked questions

Wildcards: * any sequence, ? a single character.
Letters are kept distinct. In the Vate alphabet ç ê î ş û are letters in their own right: kerdene (infinitive) and kerdêne (past continuous) are different words, and a search will not return them together. Only case and non-alphabet letters (ı ö ü ğ â é) are folded. To fold circumflexed letters as well, tick the “ê=e folding” box.
Regular expressions: enclose the query in /…/ — /ene$/ (infinitive), /^nê/ (negative), /.{3,}anê$/ (plural ezafe). Flags: i case-insensitive, d diacritic-insensitive (/kerd/d). Inside a regex, ê î û ş ç are written as they are. Pattern guide →

Search over tag sequences. Each bracket is one token; an empty [] matches any token. Fields: form, lemma, upos, xpos, feats, src. = equals, ~ contains, ! negates, | separates alternatives, /…/ is a regular expression.
Regex examples: [form=/ene$/ upos=VERB] verbs in the infinitive, [feats=/Case=Obl/][upos=VERB] oblique noun + verb.
Letters are kept distinct: [form=kerdene] and [form=kerdêne] give different results — in the Vate alphabet ç ê î ş û are separate letters. To fold circumflexes, use the regex flag d: [form=/kerdene/d].
Put a regex containing ] in double quotes: [form="/[iî]ş$/"]. Pattern guide →

Syntax: /body/flags · ^ start of word, $ end of word, . any letter, .{3,} at least three letters, [iî] alternatives, (a|b) group, (?:…) non-capturing group, \1 backreference. Flags: i case-insensitive, d diacritic-insensitive.

77 patterns. Each pattern is a regular expression and can be typed straight into the search box. The patterns were selected by measuring them against the corpus: the token counts and examples below come from the current annotated state of the corpus. The precision column gives the share of the pattern's proposed part of speech across all matching tokens: a pattern above 95% can be used on its own, one below 60% should always be combined with a part-of-speech filter. Low precision is not always the pattern's fault — some affixes really are homographic (-a ezafe vs. -a feminine participle), while in other cases the annotation is to blame (the prohibitive me- pattern).

First an affix inventory was derived from form–lemma pairs (how many types and tokens each affix covers), then every candidate pattern was measured against the corpus and its dominant part-of-speech share examined. Only patterns that discriminate well and correspond to a grammatical unit were kept; stem-length bounds (.{3,}) were added to prevent false matches on short words.

Patterns with a low dominant-tag share (e.g. preverbs) were kept to be used together with a part-of-speech filter. A pattern concentrating in an unexpected tag usually points to an annotation error; this is why pattern search doubles as an auditing tool.

Per-million frequency (pmw) is computed over the whole corpus. The filter field accepts wildcards (ker*) and regular expressions (/[iî]ş$/), so you can list every form of an affix in frequency order.
ç ê î ş û count as separate letters: kerdene and kerdêne are separate rows.

The node field accepts wildcards (kerd*) and regular expressions (/ene$/), so you can treat every form of an affix as a single node and measure its collocates together. ç ê î ş û count as separate letters.

The texts in this corpus are used with the permission of the publisher and the authors. That permission is conditional on the texts not being read or downloaded by third parties: search results are shown with limited context (±10 tokens), and full-text view and text download are disabled. Quantitative output such as frequency and collocation data may be used freely.

The texts were included in the corpus by permission of the publisher; that permission is conditional on the texts not being read or downloaded by third parties. Full-text view is therefore open only to the annotation team (editor, reviewer, administrator). The corpus as a whole is open to everyone through the concordance, frequency and collocation tools.

The corpus is automatically annotated (silver standard). 0.2% of tokens have no dictionary match yet and are awaiting human review.

If you would like to contribute, register; once an administrator grants you the contributor role you can propose corrections from the annotation panel.

This page shows two things separately: the version of the running source and the state of the database. If both reach the expected values after an update, the update has been installed. If only one is stale: a stale source means guncelle.bat could not unpack the zip; a stale database means the migration or counting scripts did not run.

Order of work

1. Pick a token from the list on the left. 2) Read the sentence. 3) If one of the candidates under Decision support is right, press “apply” — it writes lemma, part of speech and features together. 4) If not, fill the fields by hand. 5) Once you are satisfied it is correct, press “mark MANUAL / 1.0”. 6) If the decision also holds for identically spelled tokens, choose a propagation scope.

Sections

Context → where you make the decision. Decision support → what three sources propose. Identity → form and lemma. Part of speech → UPOS. Grammar → features (FEATS). Meaning → gloss. Provenance → source, confidence, approval. Bulk action → propagation. The ? in each section's corner explains what it is for and where mistakes are made.

What to approve

The MANUAL mark is permanent: that token will never again be changed by a model, a bulk correction or a propagation. Rather than approving a token you are unsure about, leave it as it is — a wrong approval damages both the accuracy measurement and the model.

Where to start

The ready-made filters open the places with the highest return: “unresolved” are tokens with no tag at all, “low-confidence model tags” are the ones the model was unsure about, “auto-propagated” are the ones awaiting verification. Filtering by a morphological pattern is another route: it lets you review every form of one affix together.

What it is for

Part of speech is read from the context, not from the word itself. The same spelling can take a different tag depending on the sentence: “newe” on its own is a noun (newness), after an ezafe-marked noun it is an adjective (çîyêko newe).

How to use it

Click any token in the sentence to move to it; the ‹ › buttons on the left step through them in order. Tokens shown in a faded colour are still unresolved (X).

Watch out

Do not decide on a part of speech without reading the whole sentence; copulas (o, a, yo, ya) and fused adpositions in particular can only be told apart in context.

What it is for

It shows side by side what three independent sources say about this form: how the form is tagged elsewhere in the corpus, what the Vate dictionary says, and what the affix analyser finds. Applying a suggestion writes lemma, part of speech and features together.

What the score means

Corpus evidence 0.95 · exact dictionary match 0.85 · affix analysis 0.70 · dictionary match differing only in diacritics 0.60. If several sources agree the score rises slightly; non-standard spellings lower it.

Watch out

A suggestion labelled “dictionary (diacritic difference)” means this exact spelling is not in the dictionary — as when a query for kerdêne returns the entry kerdene. These differ in circumflexed letters and may genuinely be different words.

What it is for

The form is whatever the text says — it is the corpus's source material. The lemma is the citation form that gathers every inflection under one entry: search, frequency and dictionary work all rest on it.

How to write the lemma

For verbs, the infinitive: kerdêne → kerdene. For nouns, the bare singular: serranê → serre. For verbal nouns, the form itself: kerdişî → kerdiş. For proper nouns, initial capital and bare stem: Dêrsimî → Dêrsim. Adverbs and adpositions stay as they are.

Watch out

Touch the form field only for a tokenisation error — changing the text changes the corpus. Keep circumflexed letters in the lemma: ç ê î ş û are separate letters in the Vate alphabet, and kerdene and kerdêne are different words.

What it is for

The backbone of every measurement: search, frequency lists, collocation, pattern search and model training all look at the part of speech. A wrong tag damages all of these layers at once.

How to decide

Look at the word's function in this sentence, not at the parts of speech it could have in the dictionary. A short definition and example of the tag you select appears under the box.

Hard distinctions

NOUN / PROPN: place and person names are PROPN (Dêrsim, Ehmed), common nouns NOUN (dewe). · ADJ / NOUN: a modifier following an ezafe-marked noun is ADJ (çîyêko newe), the same word used alone is NOUN. · VERB / AUX: copulas and auxiliaries (o, a, yo, ya, bî, est) are AUX. · DET / PRON: DET if a noun follows (na dewe), PRON if it stands for the noun (na). · PART is only for function words (ne, ha, wa); if you are unsure, do not choose PART.

Language-specific tag (XPOS)

A language-specific tag inherited from the old pipeline. Because it contradicted the part of speech, it is left empty in the new corpus and is not written on export. You are not expected to fill it in by hand.

What it is for

This is where the distinctive structures of Zazaki are recorded: grammatical gender, ezafe, oblique case, the ergative past, person–tense inflection. The FEATS patterns in pattern search (Case=Obl, Ezafe=Yes, Voice=Erg) scan this field.

How to fill it in

Only the features defined for the part of speech you selected are shown. Each value carries a gloss and an example from the corpus. For instance: keyeyî → Case=Obl, Gender=Masc, Number=Sing. In the phrase kitabê lajekî the first word carries Ezafe=Yes.

Watch out

Do not guess at a feature you cannot read from the text — leaving it empty is better than filling it in wrongly, because a wrong feature damages both the measurement and the model. Features that are not defined for the tag but already present are preserved; the raw feature-string field is for advanced corrections only.

What it is for

The gloss is the corpus's opening onto dictionary work and Leipzig IGT (interlinear glossed text); it is carried in the MISC field on export.

How to write it

Write the dictionary meaning, not the inflected one: for kerdî write “to do”, not “did”. Separate several equivalents with commas (suffix, clitic). The Turkish field is required, the English field optional.

Watch out

Punctuation, numerals and proper nouns need no gloss. Bulk glossing is written only for words that have a dictionary equivalent; the rest stay empty, and a gloss you write by hand is never changed by a bulk operation.

What it is for

It records where each tag came from. Tracing the source of an error, measuring accuracy layer by layer and undoing bulk operations all depend on it — and the information is carried on export too (MISC: Src, Conf).

The layers

RULE rule-based pipeline · MODEL trained tagger · CLAUDE hand-made part-of-speech decisions · EKKURAL affix-based correction rule · PARADIGM paradigm match · AUTO propagation from the panel (unverified) · MANUAL human approval.

How to use it

Once you have reviewed a token and decided it is correct, press “mark MANUAL / 1.0”: from then on that tag will not be changed by any bulk operation, any model or any propagation. You rarely need to edit the source and confidence fields by hand.

Watch out

The MANUAL mark is a commitment: the gold set and the accuracy measurements trust these tags. Rather than approving a token you are unsure about, leave it as it is.

What it is for

It applies your tag to the other identically spelled tokens as well. Instead of walking through millions of tokens one by one, you decide a form once and propagate that decision across the corpus.

Choosing the scope

“unresolved” is the safest: it touches only tokens tagged X. “weak layers” also takes machine tags whose confidence is below the threshold. “all but manually approved” is the broadest — it replaces every machine tag of that form with your decision. The number above each button says in advance how many tokens will be touched.

Safety

No scope touches manually approved (MANUAL) tags, hand-made part-of-speech decisions (CLAUDE, EKKURAL), or tokens that already carry an applied correction. What is written is marked AUTO / 0.6, i.e. not counted as verified; it can be undone in one step from the “Automatic tags” page, and every token returns to its previous tag.

Watch out

If a form can have more than one reading, propagation starts switched off — as with serre (year) / serre (head); unless you tick the box deliberately, the server refuses to propagate as well. Propagating across case differences is a separate checkbox and should be left off for proper nouns and abbreviations.