Appendix D — Glossary

This glossary defines technical terms used throughout the book. Entries are listed alphabetically. See: names the chapter or section where a term is explained in depth; See also: points to related entries. Gloss abbreviations come first.

Gloss Abbreviations

Interlinear glosses follow the Leipzig Glossing Rules: hyphens separate morphs and their glosses, and a dot joins several meanings expressed by one form (POSS.3SG). Numbers mark person (1, 2, 3).

Table D.1: Abbreviations used in the book’s interlinear glosses.

Label	Meaning	Label	Meaning
ABIL	ability	LOC	locative
ABL	ablative	NEG	negation
ACC	accusative	NMLZ	nominaliser
ATTR	attributive (Kazakh -ГІ, Uyghur -diki)	NOM	nominative
AUX	auxiliary	PASS	passive
CAUS	causative	PL	plural
COM	comitative	POSS	possessive
COP	copula	PROG	progressive
CVB	converb	PRS	present
DAT	dative	PST	past
GEN	genitive	PTCP	participle
INS	instrumental	Q	question particle
		SG	singular

A

Ablative case: A grammatical case marking a noun as the source, origin, or point of departure of an action (e.g., “from the city”). In Turkic languages the ablative also encodes comparison (“bigger than the city”). It is marked with -dan/-den/-tan/-ten in Turkish and -дан/-ден, -тан/-тен, -нан/-нен in Kazakh. See also: Case (grammatical), Dative case.
Accusative case: A grammatical case that marks the direct object of a transitive verb. Many Turkic languages have differential object marking. In Turkish, specific objects, including specific indefinites, usually take an overt accusative suffix (-(y)ı/i/u/ü), while non-specific objects usually appear unmarked. The conditions vary across languages (Section 5.4). See also: Case (grammatical), Nominative case.
Agglutination: A morphological strategy in which words are formed by appending a sequence of largely segmentable morphemes, often with identifiable grammatical or derivational functions. Surface shapes can alternate, and form–function mappings are not always one-to-one. Turkic languages are strongly agglutinative: a single verb form may carry tense, aspect, mood, and person agreement through a chain of suffixes. See also: Concatenative morphology, Productive morphology, Suffix template.
Allomorph: One of two or more phonological variants of the same morpheme. In Turkic languages, allomorphy is principally driven by vowel harmony and consonant assimilation. For example, the Turkish plural morpheme has the allomorphs -lar and -ler depending on the backness of the preceding vowel. See also: Vowel harmony, Morpheme.
Adapter: A small set of trainable parameters inserted into a frozen pretrained model and trained for a task, language or script. Bottleneck adapters and LoRA are different parameterisations; report the trainable modules and save the adapter with its task head. See: Section 16.4.
Annotation projection: Transferring labels from annotated source sentences to aligned target sentences through word alignments. The result is silver annotation; measure it against gold labels on the same target sentences and report unaligned words. See: Section 8.5.
Annotator agreement (Kappa): A statistical measure of the degree to which two or more human annotators assign the same label to the same item, corrected for chance agreement. Cohen’s kappa and Fleiss’s kappa are the most common variants. Interpretation depends on the task, label prevalence and chance model; report raw agreement and label counts as well. There is no universal adequacy threshold. See also: Inter-annotator agreement.
Attention mechanism: A mechanism that combines information from permitted positions using weights computed from their compatibility with a query. Self-attention relates positions within one sequence; cross-attention draws on another sequence, such as encoder outputs. The transformer uses scaled dot-product attention to combine value vectors, with masks controlling which positions are available (Vaswani et al. 2017). See also: Multi-head attention, Encoder-decoder.

B

Back-translation: A data augmentation technique for machine translation in which a target-language monolingual corpus is translated into the source language by a reverse MT system, generating synthetic source-target sentence pairs for training. Its value for a Turkic pair is an empirical question that depends on the reverse system’s quality (Chapter 10). See also: Transfer learning.
BIO tagging: A sequence labelling scheme used for span-level tasks such as named entity recognition. Each token is assigned a tag indicating whether it is at the Beginning of a span (B-), Inside a span (I-), or Outside any span (O). Extended variants include BIOES (End, Single) and BILOU. See also: Named entity recognition (NER), POS tagging.
BLEU: A machine-translation metric combining modified token n-gram precision with a brevity penalty against reference translations. It is sensitive to tokenisation and to alternative word forms, which matters for agglutinative languages; report it with its SacreBLEU signature. See also: chrF. See: Section 10.3.
Byte-Pair Encoding (BPE): A subword tokenisation algorithm that iteratively merges the most frequent adjacent pair of symbols in the training representation (characters or bytes, depending on the tokeniser), producing a vocabulary of variable-length subword units. Originally proposed for data compression and adapted for NLP by Sennrich, Haddow, and Birch (2016). BPE is used by many Turkic language models and MT systems. See also: Subword tokenisation, Fertility (tokeniser), SentencePiece.

C

Calibration: The correspondence between predicted confidence and observed outcome frequency under a specified evaluation distribution. Low entropy is not evidence of good calibration.
CARE principles: Collective benefit, Authority to control, Responsibility and Ethics: principles for Indigenous data governance (Carroll et al. 2020). They pose questions for collaboration with language communities rather than a checklist that settles them. See: Section 17.4.
Case (grammatical): A grammatical category that encodes the syntactic or semantic role of a noun phrase in a clause. Turkic languages use suffixal case marking; case inventories depend on the language and grammatical analysis; see Chapter 5. See also: Nominative case, Accusative case, Dative case, Genitive case, Ablative case.
chrF: A machine-translation metric based on character n-gram overlap with reference translations; chrF++ adds word n-grams. It is less sensitive than BLEU to suffix variation, but a meaning-changing suffix such as negation can leave it almost unchanged. See also: BLEU. See: Section 10.3 and Section 15.2.
Code-switching: The alternation between two or more languages within a conversation, sentence, or word, e.g. Kazakh–Russian or Turkish–German mixing. See: Section 16.6.
Common Turkic Alphabet and Common Turkic Script (CTS): A 34-letter Latin alphabet proposed in 1991; a revised version was agreed in 2024 by the Turkic World Common Alphabet Commission as a recommendation to the member states. In this book, CTS names TurkicNLP’s versioned common-script mapping, used as a shared representation for text processing. See: Section 3.6.
Complementiser: A functional morpheme that introduces a complement clause. Turkic languages often express complementiser-like functions through nominalised or participial verb forms. Finite subordination also uses elements such as Turkish ki and diye.
Concatenative morphology: Building forms by joining morphemes in a linear sequence. Their surface shapes may alternate. Turkic languages rely heavily on suffixation, alongside alternations, reduplication and other processes. See also: Agglutination.
CoNLL-U: The tab-separated file format of Universal Dependencies: one word per line with ID, form, lemma, UPOS, XPOS, features, head, relation, enhanced dependencies and miscellaneous fields. Range IDs mark multiword tokens; decimal-style IDs mark empty nodes in enhanced dependencies. See: Section 8.2.
Contamination (benchmark): Evaluation items, or close variants of them, appearing in a model’s training data, so that a score can reflect exposure rather than ability. High scores are not proof of it; an independent check is needed. See: Section 15.3. See also: Data leakage.
Contextual embedding: A vector representation of a word or token that is computed in the context of the surrounding sentence, as opposed to a fixed lookup-table embedding. Contextual embeddings are the output of each layer of a transformer encoder; they capture disambiguation information that static word vectors cannot represent. See also: Masked language model (MLM), Fine-tuning.
Converb: A non-finite verb form used in adverbial clauses, for example to express time or manner. Turkic converbs also occur in clause chains and auxiliary constructions; their syntax depends on the form and construction. See: Section 2.5.
Cross-entropy: A loss function used to train classification and language models, measuring the expected negative log probability assigned by a model \(Q\) to outcomes distributed as \(P\): \(H(P,Q)=H(P)+D_{KL}(P\|Q)\). It is not itself a divergence. In language modelling, cross-entropy over a test set can be converted to perplexity by exponentiation with the matching log base (Section A.2).

D

Data leakage: Test data influencing model training, fitted preprocessing or model selection, so that the test evaluation is no longer independent. Examples include duplicates across splits, translated test items in training, and fitting a tokeniser on the test corpus. Applying a fixed preprocessing rule to test input is ordinary inference.
Dative case: A grammatical case marking the indirect object or goal of an action (e.g., “to the city”). Turkic dative suffixes take the form -(y)a/-(y)e in Turkish and -ға/-ге/-қа/-ке in Kazakh, with allomorphs selected by vowel harmony and by the stem-final consonant. See also: Case (grammatical), Ablative case.
Dependency parsing: A syntactic analysis task that assigns a head word to each word in a sentence and labels the directed arc between them with a grammatical relation (subject, object, etc.). The Universal Dependencies framework provides a cross-linguistically consistent annotation scheme used across the Turkic UD treebanks. See: Chapter 8.
Derivational morphology: The branch of morphology concerned with forming new lexical items from existing ones, typically by adding derivational affixes that change the word’s part of speech or core meaning (e.g., Turkish göz ‘eye’ → göz-lük ‘glasses’). Turkic languages have extensive derivational morphology. See also: Inflectional morphology, Agglutination.
Disambiguation (morphological): The task of selecting the correct morphological analysis for a word form given its sentential context, when multiple analyses are possible. Agglutinative languages like Turkish and Kazakh frequently produce highly ambiguous surface forms, making disambiguation a non-trivial component of NLP pipelines. See also: Morphological analysis, FST (finite-state transducer).
Distribution shift: A mismatch between the statistical distribution of the data used to train a model and the data it is applied to at inference time. For Turkic NLP, distribution shift is common when models trained on literary or news text are applied to social media, SMS, or spoken transcripts.
Dropout: A regularisation technique for neural networks in which a random subset of activations is set to zero during training. It can reduce co-adaptation between units; its effect on generalisation must be evaluated.

E

Embedding: A learned vector representation of a token, word, sentence or other object. Vectors of equal dimension need not be semantically aligned across languages.
Encoder-decoder: A neural architecture consisting of an encoder that maps an input sequence to a latent representation and a decoder that generates an output sequence from that representation. The architecture is standard for sequence-to-sequence tasks including machine translation, text summarisation, and morphological generation. See also: Attention mechanism, Masked language model (MLM).
Evidential: A grammatical category marking the source of information, such as indirect or reported knowledge. Turkish -mIş has evidential readings (UD Evident=Nfh); translation must choose how to render them. See: Section 5.1.

F

Fertility (tokeniser): The number of subword tokens divided by the number of words under a declared word definition and tokenisation protocol. Corpus-total ratios differ from averages of sentence ratios. Report whether words were encoded separately or as running text, and whether special tokens count. High fertility describes fragmentation; neither its cause nor the benefit of script conversion follows from it. See also: Byte-Pair Encoding (BPE), Subword tokenisation.
Fine-tuning: The process of continuing the training of a pretrained model on additional data for a task or language objective, with or without manual labels, updating all or a subset of model parameters. Fine-tuning is the dominant paradigm for adapting BERT-class models to downstream Turkic NLP tasks. See also: Transfer learning, LoRA, Parameter-efficient fine-tuning (PEFT).
FST (finite-state transducer): A computational device that maps one string to another through a finite set of states and transitions, with transitions that may consume or emit epsilon (the empty string). It can define multiple outputs per input and attach weights to paths. FSTs are the core technology of rule-based Turkic analysers such as the Apertium modules and TRmorph. See: Chapter 6.

G

Genitive case: A grammatical case marking possession or modification. In Turkic languages, the genitive-possessive construction is a two-part structure: the possessor noun takes a genitive suffix and the possessed noun takes a possessive agreement suffix cross-referencing the possessor’s person and number. See also: Case (grammatical).
GlotLID: An open language identifier with more than 2,000 language-and-script labels, used by TurkicNLP (Kargaran et al. 2023). It can confuse closely related languages, especially in short texts. Evaluate its accuracy for the intended languages and input lengths. See: Section 3.7.

I

Inflectional morphology: The branch of morphology concerned with grammatical forms of a lexical item, expressing categories such as tense, aspect, mood, number, person and case. The boundary between inflection and derivation depends partly on the analysis. Turkic languages have rich inflectional paradigms, particularly on verbs and nouns. See also: Derivational morphology, Agglutination.
Inter-annotator agreement: The degree to which independent annotators produce consistent labels for the same data. See: Annotator agreement (Kappa).

L

Labial harmony (rounding harmony): A subtype of vowel harmony in which the feature [round] spreads from a vowel to subsequent vowels in the word. Turkish rounds high suffix vowels; Kyrgyz also rounds low vowels and writes it, while Kazakh rounding is pronounced but not written (Section 2.2). See also: Vowel harmony, Palatal harmony.
Language identification (LID): Automatically determining the language of a text. Closely related Turkic languages written in the same script are frequently confused with each other. See: Section 3.7.
Lemma: The canonical dictionary form of a word, abstracting away from inflectional variation. The citation convention depends on the resource: a verbal stem and an infinitive may be different lemma conventions. Match the dataset’s scheme before scoring. See also: Morphological analysis, Inflectional morphology.
Light-verb construction: A complex predicate formed by a nominal and a semantically light verb, such as Turkish yardım et- ‘help’ (lit. ‘do help’). UD labels the relation compound:lvc, but released Turkic treebanks differ in which word is the head. See: Section 8.2.
LoRA (Low-Rank Adaptation): A parameter-efficient fine-tuning method that learns an additive weight update as the product of two smaller matrices while keeping the original weight matrix frozen. The chosen rank and adapted modules determine the number of trainable parameters. QLoRA combines it with a quantised base model. The book’s LoRA results are in Chapter 12. See also: Parameter-efficient fine-tuning (PEFT), Fine-tuning.

M

Macro-F1: The unweighted mean of per-class F1 scores, so that small classes count as much as large ones; micro-F1 pools all decisions instead. State which one is reported, and for NER whether spans are scored strictly. See: Section 9.2.
Masked language model (MLM): A pretraining objective in which selected input positions are prediction targets; depending on the recipe, they may be masked, replaced or left unchanged. The model predicts their original tokens. MLM is the pretraining objective of BERT-class models such as BERTurk, XLM-R and Glot500 (Chapter 12). See also: Contextual embedding, Fine-tuning.
Morpheme: The minimal unit of meaning or grammatical function. Roots and affixes are both morphemes. Turkic affixation is predominantly suffixal, and a single word may contain five or more morphemes. See also: Agglutination, Allomorph.
Morphological analysis: Identifying a word form’s lemma, part of speech and morphological features, often with alternative readings. Some analysers also provide morpheme segmentation; a UD feature bundle does not itself specify morpheme boundaries. Explicit analysis can support downstream tasks, but end-to-end models need not call a separate analyser. See also: FST (finite-state transducer), Disambiguation (morphological).
Morphological generation: The inverse of morphological analysis: producing a surface word form from a lemma and a specification of morphological features. An explicit generator is useful in rule-based MT and controlled form production; neural MT and TTS need not contain a separate morphological generator. See also: Encoder-decoder, FST (finite-state transducer).
Multi-head attention: An extension of the attention mechanism in which multiple attention functions are computed in parallel on linearly projected versions of the input, allowing the model to attend to different aspects of the input simultaneously. The transformer’s self-attention layers use multi-head attention. See also: Attention mechanism.

N

Named entity recognition (NER): A sequence labelling task that identifies and classifies mentions of named entities (persons, organisations, locations, etc.) in text. Turkic NER resources include KazNERD (Kazakh, 25 entity classes), KyrgyzNER and WikiANN (multilingual silver-standard NER). See: Section 9.2.
Nominative case: The default, unmarked case in Turkic noun paradigms, used for subjects of sentences and predicate nominals. Because the nominative is often zero-marked, it doubles as the citation form of nouns. See also: Case (grammatical), Accusative case.

O

Out-of-vocabulary (OOV): A word or other unit absent from a declared vocabulary. A surface form can be unseen even when its lemma is familiar. Productive morphology creates many possible forms, but observed OOV rates also depend on corpus size, domain and tokenisation. Distinguish unseen word forms from unknown subword tokens (Section 15.2). See also: Subword tokenisation, Fertility (tokeniser).

P

Palatal harmony: A subtype of vowel harmony in which the feature [back] (or [front]) spreads within a defined harmonic domain. Which vowels participate and which suffixes alternate depends on the language, construction and exceptions. Palatal harmony is the most pervasive form of vowel harmony in Turkic, present in Turkish, Kazakh, Kyrgyz, and most other members of the family. See also: Vowel harmony, Labial harmony.
Parameter-efficient fine-tuning (PEFT): A family of methods for adapting pretrained language models to new tasks or languages by updating only a small subset of parameters or adding a small number of new parameters, leaving the bulk of the pretrained weights frozen. They reduce the memory and storage needed to adapt large models; Chapter 12 compares LoRA with full fine-tuning. See also: LoRA, Fine-tuning.
Participial clause: A clause built around a participial verb form, with a syntactic role that depends on the language and construction. Turkic languages commonly place participial relatives before a noun, alongside finite and other relative strategies. Participial clauses can also serve other functions. See also: Converb, Dependency parsing.
Pivot translation: Translating through an intermediate language (source → pivot → target) when a direct system or data are lacking. Each leg must be supported, and errors can accumulate or be repaired along the way; compare it with direct translation on the same test set. See: Section 10.4.
POS tagging: The task of assigning a part-of-speech label to each token in a text. Universal Dependencies defines a set of 17 universal POS tags (UPOS) used across all Turkic treebanks. See also: UPOS (Universal POS), Dependency parsing, BIO tagging.
Pro-drop: A grammatical property of languages in which overt pronoun subjects are omitted when they can be recovered from verb morphology or context. Null subjects are common in Turkic; their licensing depends on grammar and discourse. Person/number agreement can help identify a subject, but Turkic subject agreement does not encode grammatical gender. See also: SOV.
Productive morphology: Morphological processes that apply freely to new or nonce words, generating valid forms according to the rules of the language. Turkic inflection and derivation can produce many forms absent from a training corpus. Productivity varies by process and does not mean that any suffix can attach to any stem. See also: Agglutination, Inflectional morphology.
Proxy language: A language selected as a model routing code, language embedding or transfer source for a target with limited task supervision. This operational choice does not reveal an internal representation learned by the model. See: Chapter 16; also Chapter 7 and Chapter 8.

R

Retrieval-augmented generation (RAG): Generation conditioned on retrieved evidence. Evaluate retrieval relevance, answer correctness, citation support, language fidelity and abstention separately.
Romanisation: The representation of a text written in a non-Latin script using Latin characters. For Turkic NLP, transliteration arises in Cyrillic-to-Latin conversion (Kazakh, Uzbek) and Arabic-to-Latin conversion (Uyghur, Ottoman Turkish). See: Section 3.4.
Round trip (transliteration): Converting text into another script and back and checking whether the original is recovered. It tests reversibility, not correctness of the intermediate text: a lossy mapping fails it, and a consistent error can pass it. See: Section 3.4.

S

Scrambling (word order): The syntactic phenomenon in which arguments of a verb can appear in multiple linear orders without changing the basic propositional meaning, though information structure (topic, focus) typically changes. Turkic languages allow considerable word-order variation around an SOV base; this flexibility complicates dependency parsing and semantic role labelling. See also: SOV, Pro-drop.
SentencePiece: A tokeniser framework supporting BPE and unigram language-model algorithms. It is not a synonym for the unigram algorithm. Record the algorithm and normalisation configuration.
Silver data: Automatically produced annotation, for example from a tagger or annotation projection. It may supplement manually checked gold data or be used where gold training data are unavailable. See: Section 8.5.
SOV: Subject-Object-Verb: a typological label for languages in which the canonical order of major clause constituents places the subject first, the object second, and the verb last. SOV is the usual base order in Turkic, with contact-induced exceptions as well as discourse and construction effects (Section 8.1). See also: Scrambling (word order).
Subword tokenisation: A tokenisation strategy whose vocabulary can contain whole words, word parts and individual characters or bytes, depending on the method. BPE, WordPiece and unigram language-model tokenisation can represent unseen words as sequences of known units. Unknown-token behaviour depends on character coverage and fallback rules. Subword tokenisation does not change whether a whole word was seen in training (Section 15.2). See also: Byte-Pair Encoding (BPE), Fertility (tokeniser), Out-of-vocabulary (OOV).
Suffix template: The ordered slot-structure specifying the sequence of morphological categories that can be realised as suffixes on a word stem. A common Turkic verbal order is stem → derivation → voice → ability → negation → tense/aspect/mood → person agreement (konuş-tur-a-ma-dı ‘could not make speak’, Section 2.1). Templates are construction-specific generalisations, not fixed slots in every word (Section 5.1). See also: Agglutination, FST (finite-state transducer).

T

Temperature sampling: Sampling training languages with probability proportional to \(n_i^{1/T}\), where \(n_i\) is the data size of language \(i\). \(T=1\) follows the data sizes, larger \(T\) moves towards uniform; small languages then repeat their few examples more often. See: Section 16.4.
Transfer learning: A machine learning paradigm in which a model trained on one task or dataset is adapted to a different (often related) task or dataset, exploiting shared structure. In Turkic NLP, cross-lingual transfer from well-resourced languages (Turkish, Kazakh) to lower-resourced relatives (Kyrgyz, Uzbek) is a primary strategy for building systems with limited annotated data. See also: Fine-tuning, Zero-shot transfer.
Transliteration: Conversion between writing systems under a declared mapping. Romanisation specifically targets Latin script. Mappings may be lossy or context-sensitive; preserve originals and offsets.
Treebank: A corpus of sentences annotated with syntactic structure, such as constituency trees or dependency graphs. The release-specific inventory in Chapter 8 lists the Turkic UD resources, including the UD-Turkic parallel treebanks. See also: Dependency parsing, UPOS (Universal POS).

U

UPOS (Universal POS): The set of 17 coarse-grained, cross-linguistically applicable part-of-speech tags defined by the Universal Dependencies framework: NOUN, VERB, ADJ, ADV, PRON, DET, ADP, AUX, CCONJ, SCONJ, PART, INTJ, NUM, PUNCT, SYM, X, and PROPN. See also: POS tagging, Treebank.

V

Vowel harmony: A phonological process in which participating vowels within a harmonic domain are constrained by one or more phonological features, most commonly frontness/backness (palatal harmony) and/or rounding (labial harmony). Harmony is widespread in Turkic, but its domains, strength, exceptions and spelling vary; standard written Uzbek has largely fixed suffix vowels (Section 2.2). See also: Palatal harmony, Labial harmony, Allomorph.

Z

Zero-shot transfer: The application of a model trained on one language (or task) to a different language (or task) without any task-specific fine-tuning on the target language. The model may still have seen target-language text during pretraining (Section 16.1); zero-shot here means no target task-labelled training. See also: Transfer learning, Fine-tuning.