Mova0.6 pre-release

the flag of Kazakhstan 34. Kazakh in Mova

The Kazakh package uses the language code kk. Mova’s multilingual pipeline distinguishes language identification from grammatical analysis. Standard Kazakh Cyrillic currently falls into the segmenter’s other category, and the package has neither a parser model nor an annotated corpus. The distinction between missing output and measured performance matters when reading the package status.

1. Language segmentation and Cyrillic input

Level 1 of tools/mova-parse uses tools/langseg to assign a language label to each sentence and span. A span is the stretch of input that level 2 can send to a parser for the language identified there. The label therefore determines which language model can be selected for that span.

The segmenter distinguishes English and French in Latin script using a Latin character table, a lexicon for each language and a character 4-gram model. It also distinguishes Hebrew-script varieties and recognizes Hebrew or Aramaic in scholarly transliteration. The remaining scripts fall under other. Kazakh in standard Cyrillic falls under that last category, so the label indicates that the script is outside the segmenter’s supported set. It supplies no identification of the span as kk.

Consider a sentence in standard Kazakh Cyrillic (constructed).

FormRomanizationGloss
БалаBalachild
кітапkitapbook
оқыдыoqy-dyread-PST

Translation: “The child read a book.”

The Cyrillic letters in the example allow a reader to recognize the script, and the words allow a Kazakh reader to understand the sentence. The segmenter’s script coverage gives it only the other label for this input. In a text that also contains a stretch of supported Latin-script material, the segmenter can assign a supported language to that stretch independently of the label on the Cyrillic material.

A further Latin-script language can join the segmenter by adding a lexicon and a Lang code. The lexicon can be built from open Universal Dependencies (UD) training data with langseg latin-lexicon. This procedure concerns Latin input. The romanization column above is a reading aid, and presenting the example in that form gives the reader no basis for expecting a Kazakh label from the current segmenter.

2. Missing Kazakh model and Parse=None output

Level 2 of tools/mova-parse selects a model for the language of each span. Its selection is language-specific. English and French have model files, en.bin and fr.bin, respectively. Kazakh has no level-2 model. The absence of Kazakh recognition for standard Cyrillic is a level-1 limitation, while the absence of a model is a separate condition at level 2.

When the parser has no model for a language, the # parse comment reports the missing model. In the unparsed output, the fields in the table have the following values. Kazakh token rows carry this unannotated output, including the Parse=None marker.

FieldValue for an unparsed tokenAnnotation that is unavailable
LEMMA_Dictionary form
UPOS_Universal part-of-speech tag
XPOS_Language-specific part-of-speech tag
FEATS_Morphological features
HEAD_Index of the syntactic head
DEPREL_Dependency relation to the head
MISCParse=None markerThe marker records the absence of a parse

The comment describes the missing model at the parse level, while the marker makes the status visible on individual rows. A row carrying the marker provides no evidence about how Kazakh grammar analyzes the displayed form. For example, an underscore in FEATS supplies no judgment about whether a nominal form has case, and an underscore in HEAD supplies no judgment about whether a word attaches to the predicate or serves as the root. Likewise, the presence of token rows in the output does not establish that Mova has a Kazakh tokenizer. Kazakh tokenization, tagging, lemmatization and dependency parsing are all unavailable.

3. Annotated data held by Mova and the unmeasured baseline

Mova holds no annotated Kazakh corpus. There is therefore no measured Kazakh treebank baseline in the package. This inventory concerns data held by Mova. It gives no account of what Kazakh resources may be available elsewhere.

An annotated treebank supplies reference analyses for training and evaluation. A training split provides sentences paired with annotations that the tagger and parser can learn from. A development split provides separate reference trees against which to compare the resulting predictions. A grammar example, even one with a morpheme-by-morpheme gloss, provides only the analysis given in that example. A gloss does not supply the full tokenization, tags, head indices and dependency labels needed to evaluate a UD tree.

ItemKazakh package status
Annotated corpus held by MovaNone
Level-2 parser modelUnavailable
Development-set unlabeled attachment score (UAS)Unmeasured
Development-set labeled attachment score (LAS)Unmeasured

An unmeasured score is missing data. The number of rows marked Parse=None describes missing analyses in an output, while UAS and LAS compare predicted attachments with annotated reference attachments. Reading the marker as an attachment error would conflate these two observations and would turn a gap in coverage into a purported evaluation result.

4. The UD training and evaluation workflow

The language package specification in What every language gets calls for the same train-to-development baseline for every language. Mova’s parser is in the en crate, which supports training on a language’s own UD training split.

StageCommandRole in the baseline
Prepare training dataen xpos-fillPrecedes parser and tagger training on the UD training split
Trainen trainFits the tree parser and tagger to the language’s training data
Evaluateen parse-evalMeasures UAS and LAS on the development split after training

UAS gives credit when the predicted head matches the reference head, regardless of the dependency label. LAS requires both the head and the dependency label to match for the same attachment. These two scores describe dependency prediction. They do not measure how often the segmenter identifies Kazakh input or how often downstream components recover the meaning of an utterance.

Consider a sentence that distinguishes the information needed for a head decision from the information needed for a labeled attachment (constructed).

FormRomanizationGloss
ОлOl3SG
мектепкеmektep-keschool-DAT
бардыbar-dygo-PST

Translation: “They went to school.” The subject is singular.

In a reference analysis that attaches мектепке to барды as an oblique dependent, an attachment to барды gets credit for UAS. To get credit for LAS, the prediction also needs the reference label obl. The gloss DAT describes the noun’s case marking. The dependency label describes the syntactic role of the nominal with respect to its head. Each belongs to a different part of the annotation.

The workflow needs an annotated training split to produce a model and an annotated development split to measure attachment. Running a trained model on raw text through the multilingual parser also requires that the segmenter identify the language of the span so the parser can select the model. The train-to-development procedure specifies how a Kazakh baseline would be measured once the data are available. There is no current result for kk.

5. Current limits on frames and pragmatics

Mova’s frame, pragmatic and world components, including prag and world, read English trees. Their input is a syntactic analysis. Sentence and span labels from language segmentation do not provide the predicate and dependency structure that a tree supplies. Kazakh rows marked Parse=None therefore provide no analyzed tree for these components to read.

In the example with мектепке, a reader can understand the school as the destination. The gloss identifies dative case, and the translation conveys motion toward that destination. Neither is Mova output. The missing Kazakh parse means there is no syntactic analysis of this sentence for downstream processing, so there is no Kazakh frame or pragmatic interpretation to report.

A Kazakh tree is a prerequisite for extending this processing to Kazakh. The documented tree input for the downstream components is English. A UD attachment score, if one becomes available for Kazakh, would report how often dependency predictions match the reference and would leave downstream interpretation unmeasured.

6. Squid claims and implementation status

Squid reads the grammar manual and turns every grammatical point into a claim. Formalizable claims, including rules, closed lists and statistics, wait in Squid’s queue for a place in Mova’s code. Queue status records that a claim is awaiting implementation. It establishes no behavior of the parser, tagger or other components.

A single construction can raise different questions for the system. The form мектепке illustrates the distinction. A spelling or morphology claim can describe the attachment of the dative suffix. A tagging claim can describe which feature represents the resulting case value. A syntactic claim can describe the conditions under which the form is attached as an oblique. The gloss and translation can help a reader check the intended reading, while each claim still needs its own conditions for application in code.

For Kazakh, the first points to address are those needed for a tokenizer, a tagger and a UD treebank baseline. Tokenization determines the units available for annotation. Tagging assigns categories and features to those units. Dependency annotation specifies their heads and relations so that predicted trees can be scored. The present Kazakh grammar claims have no implemented rules in Mova, and the baseline is still unmeasured.

7. How Mova handles Kazakh now

For standard Cyrillic input, level 1 of tools/mova-parse uses tools/langseg and assigns other. This gives Mova no Kazakh language identification for the span. At level 2, no model exists for kk. A missing model is reported in the # parse comment, and unparsed token rows carry _ in LEMMA, UPOS, XPOS, FEATS, HEAD and DEPREL, with Parse=None in MISC. Kazakh tokenization, tagging, lemmatization and parsing are unavailable.

Mova holds no annotated Kazakh corpus, so there is no measured UAS or LAS baseline. The workflow in What every language gets uses en xpos-fill before en train on the training split and en parse-eval to measure attachment on the development split. Mova’s frame, pragmatic and world components read English trees, and there is no Kazakh parse to supply downstream.

Squid reads the grammatical points as claims. Formalizable claims wait for implementation, beginning with the points needed for tokenization, tagging and a UD treebank baseline. No Kazakh grammar rule described in the manual is implemented yet.

Sources for this chapter

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/kk/34-mova-tools/