Mova0.6 pre-release

the flag of Kyrgyzstan 34. Kyrgyz in Mova

Mova’s multilingual parser has two levels. The first assigns language labels to sentences and spans. The second uses language-specific models to produce Universal Dependencies analyses. Standard written Kyrgyz uses Cyrillic, which the current segmenter assigns to other. Mova has neither a Kyrgyz parser model nor an annotated Kyrgyz corpus, so there is no measured Kyrgyz treebank baseline.

Kyrgyz Cyrillic and the current language segmenter

The first level, tools/langseg, assigns a language label to each sentence and span. For Latin script it distinguishes English and French using a Latin table, a lexicon for each language and a character four-gram model. It also distinguishes Hebrew-script varieties and recognizes Hebrew or Aramaic in scholarly transliteration. Scripts outside its coverage receive other. Kyrgyz Cyrillic falls into that category.

In the following example, the Cyrillic determines the segmenter’s outcome. The Latin romanization is included to help readers follow the gloss.

Бала китеп окуду.
Bala kitep okudu.
child book read-PST.3SG
‘The child read a book.’ (constructed)

The sentence illustrates the distinction between identifying a script and assigning a language code. The segmenter puts the Cyrillic text in other, a category shared with other scripts outside its coverage. It supplies no positive Kyrgyz identification for the sentence or its span. The pronouncing and glossing information shown here is part of the example.

There is an extension route for an additional Latin-script language. It requires a lexicon and a Lang code. The langseg latin-lexicon command builds the lexicon from an open Universal Dependencies training split. That route concerns Latin-script input. Coverage for ordinary written Kyrgyz would require the segmenter to recognize Kyrgyz in Cyrillic and assign it the language code ky.

Level-two model status and missing annotations

At the second level, tools/mova-parse selects a model for each span according to the span’s language. English and French have model files, en.bin and fr.bin. There is no level-two model for ky. Assigning the ky code to a span would enable language-based model selection, but a Kyrgyz model would still be needed to analyze it.

The multilingual parser reports a missing model in the # parse comment. In its output for Kyrgyz, the annotation fields below carry _, and MISC carries Parse=None.

FieldInformation normally recordedMissing-model value
LEMMADictionary form associated with the token_
UPOSUniversal part-of-speech tag_
XPOSLanguage-specific part-of-speech tag_
FEATSMorphological feature values_
HEADIndex of the token’s syntactic head_
DEPRELDependency relation to that head_
MISCMiscellaneous token information, including parse status hereParse=None

These values describe what the output supplies. An underscore in FEATS leaves the token’s morphological analysis unspecified. An underscore in HEAD supplies no syntactic attachment, and an underscore in DEPREL supplies no relation. Parse=None marks the absence of a parse. Token rows that carry this marker provide no evidence that Mova performed Kyrgyz tokenization or assigned Kyrgyz word boundaries.

The two levels therefore account for different gaps. The first determines whether a Cyrillic span can be identified as Kyrgyz. The second determines whether a span identified as Kyrgyz can receive lemmas, tags, features and a dependency tree. A sentence-level language label alone supplies none of those token annotations.

Corpus status and the unmeasured treebank baseline

Mova holds no annotated Kyrgyz corpus. There is consequently no Kyrgyz training and development split in the package on which to establish its requested treebank baseline. There are also no Kyrgyz corpus sentence identifiers to cite for the examples in this chapter. This status concerns the material held by Mova.

A Universal Dependencies treebank provides sentences with token boundaries and reference annotations. For a dependency baseline, each evaluated syntactic word needs a reference head and dependency relation. Part-of-speech tags, lemmas and morphological features represent additional parts of the annotation. The training split supplies examples from which a model can learn, while the development split supplies reference analyses against which the model’s predictions can be compared.

A written grammar can specify how constructions work, and a constructed example can illustrate a possible analysis. A treebank puts annotation decisions into a form usable for training and evaluation across sentences. For example, in the constructed sentence above, k itep would need a head and a relation even though its form carries no overt accusative suffix. An annotation records the word’s role in the particular sentence.

The Kyrgyz baseline is unmeasured. There is no reference run from which to infer attachment accuracy or performance on particular constructions. In particular, the missing-model marker is a processing status. It contributes no observation of how a trained Kyrgyz model would attach words or label their relations.

Training and evaluation with UAS and LAS

Mova’s tree parser and tagger are in the en crate. The package description in What every language gets calls for the same baseline procedure for each language. For a language with a Universal Dependencies training split, en xpos-fill precedes training. Then en train trains the tree parser and tagger. en parse-eval evaluates a model trained on the training split against the development split and reports UAS and LAS.

MeasureFull nameCondition for a correct attachment
UASUnlabeled attachment scoreThe predicted head is the reference head.
LASLabeled attachment scoreThe predicted head and dependency relation both match the reference annotation.

Both scores express the share of evaluated items that meet the respective condition. Consider two hypothetical errors in analyzing the example sentence. A prediction that attaches kitep to okudu and labels it nsubj in place of the reference obj gets the head right and the relation wrong. It counts as correct for UAS only. A prediction that labels kitep obj but attaches it to bala in place of okudu fails both conditions. These are illustrations of the measures using a constructed sentence.

Attachment scores describe agreement with a particular reference treebank. Lemmatization accuracy and part-of-speech tagging accuracy are separate quantities. A high attachment score by itself would supply no score for either, and an attachment baseline would not give a measurement of language segmentation. The requested train-to-development run gives the package a common way to report attachment results once the language has the required data. Kyrgyz has no result from that procedure.

Dependencies of frames, pragmatics and world knowledge

Mova’s frame, pragmatic and world components read English trees. A dependency tree records syntactic relations. A frame analysis concerns such matters as the event expressed by a predicate and the participants associated with it. Pragmatic analysis concerns how an utterance is used in context, and world knowledge can bear on how its content is interpreted.

The example sentence illustrates these levels. A dependency analysis can attach bala as subject and kitep as object of okudu. In a description of the reading event, the child is the reader and the book is what was read. The gloss makes this interpretation available to a reader of the manual. Mova’s present output for Kyrgyz supplies no tree from which its downstream components could obtain the syntactic relations.

The relevant components include prag and world. Their current input is English trees. For Kyrgyz, a parser would first need to provide trees. The existence of a Kyrgyz parser would establish that prerequisite, while downstream interpretation would still depend on support for those trees. Language identification, dependency parsing and frame or pragmatic interpretation thus concern different information about the same input.

Squid grammar claims and their implementation status

Squid reads the grammar manual and turns each 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. The queue records points available for implementation. A queued claim has no effect on parsing by virtue of being in the queue.

A grammar claim describes a linguistic pattern with its conditions and exceptions. The form of a claim can make it relevant to a particular stage of analysis. A statement about the distribution of suffix forms provides information a tagger could use when assigning morphological features. A statement about which dependents a predicate takes provides information useful when defining reference annotations. The stage at which a claim could be used depends on what its inputs and outputs are.

The constructed example demonstrates why that distinction matters. Recognizing the word kitep and finding its boundaries is a tokenization task. Identifying it as a noun is a tagging task. Attaching it as an object of okudu is a dependency task. A statement in the grammar that describes the construction can inform those tasks, but the statement itself provides no runtime result for the input sentence.

The Kyrgyz rules described in the manual are unimplemented. The first points for implementation are those needed by a tokenizer, a tagger and a Universal Dependencies treebank baseline. Tokenizer claims concern where text is divided into tokens. Tagger claims concern word classes and morphological analyses. Baseline claims concern the reference decisions needed to represent syntactic constructions as trees. Implementing those points in the relevant components is a separate step from expressing them as claims.

How Mova handles Kyrgyz now

At level one, tools/langseg assigns Kyrgyz Cyrillic to other. Its current coverage includes English and French in Latin script, Hebrew-script varieties, and Hebrew or Aramaic in scholarly transliteration. Kyrgyz Cyrillic receives no ky label from that coverage.

At level two, tools/mova-parse has no ky model. It reports a missing model in the # parse comment, leaves LEMMA, UPOS, XPOS, FEATS, HEAD and DEPREL as _, and puts Parse=None in MISC. Mova currently does no Kyrgyz tokenization, tagging, lemmatization or dependency parsing.

Mova holds no annotated Kyrgyz corpus and has no measured Kyrgyz treebank baseline. The baseline requested in What every language gets uses en xpos-fill before en train and measures UAS and LAS on the development split with en parse-eval. The existing level-two models are en.bin and fr.bin.

Mova’s frames, pragmatics and world components, including prag and world, read English trees. Kyrgyz currently supplies no parsed trees to them. Squid records the manual’s grammatical points as claims and queues formalizable claims for a place in the code. The Kyrgyz rules are still unimplemented. The first points are those needed for tokenization, tagging and the Universal Dependencies treebank baseline.

Sources for this chapter

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