Modern Uyghur Dependency Treebank

• The paper introduces MUDT, a framework with 18 main dependency relations and 26 subtypes tailored to address the limitations of universal schemes in Uyghur NLP.
• MUDT employs a four-layer morphological representation, capturing surface forms, lemmas, segmentation, and bundled features to accurately reflect complex agglutinative structures.
• Cross-standard evaluations show a 47.90% divergence from UD, underscoring the importance of refined case roles and predicate structures for improved downstream applications.

The Modern Uyghur Dependency Treebank (MUDT) is a dependency annotation framework for Modern Uyghur, introduced to address a critical resource gap in Uyghur NLP and to overcome limitations of existing treebanks for a low-resource, agglutinative language. Zuo et al. (2025) define MUDT as an integrated morphosyntactic framework grounded in typological accuracy and semantic transparency, with an inventory of 18 main dependency relations and 26 subtypes, a four-layer morphological representation, and nine annotation principles intended to capture Uyghur-specific structures that are not adequately represented by universal schemes (Zuo et al., 29 Jul 2025).

1. Definition, scope, and design objectives

MUDT is explicitly formulated as a dependency annotation framework rather than a minimal adaptation of an existing universal schema. Its scope includes 18 main dependency relations and 26 fine-grained subtypes. The framework is designed for Modern Uyghur, with particular attention to phenomena identified as problematic under existing treebank conventions, including zero copula, nuanced instrumental functions, complex predicates, and postpositional phrases (Zuo et al., 29 Jul 2025).

The central design objective is semantically transparent morphosyntactic representation. In the abstract, the framework is described as providing “a more accurate and semantically transparent representation,” and the empirical analysis is used to motivate the claim that universal schemes are inadequate for handling Uyghur-specific structures. This positioning is important because MUDT is not presented merely as an annotation variant; it is presented as a corrective response to systematic representational loss in existing resources (Zuo et al., 29 Jul 2025).

A common assumption in multilingual dependency annotation is that a universal inventory can be transferred to morphologically complex languages with limited distortion. The MUDT study directly tests that assumption through cross-standard evaluation. The reported results suggest that, for Modern Uyghur, annotation design is not a neutral formatting choice but a determinant of what syntactic and semantic distinctions remain recoverable in the treebank.

2. Morphological architecture and dependency inventory

A defining feature of MUDT is its four-layer morphological representation. The four layers are: Surface Form, Lemma, Morpheme Segmentation, and Feature Bundle (FEATS) in CoNLL-U style (Zuo et al., 29 Jul 2025). This architecture embeds morphological decomposition directly into the annotation workflow rather than treating morphology as ancillary metadata.

The framework’s example for the word “oquyanlarimdin” (“from the ones I have read”) illustrates the four layers precisely. The Surface Form is oquyanlarimdin; the Lemma is oqu; the Segmentation is oqu-yan-lar-im-din; and the FEATS are Tense=Past | Number=Plur | Person=1 | Case=ABL (Zuo et al., 29 Jul 2025). This example shows that MUDT encodes both derivationally and inflectionally relevant segmentation together with bundled morphosyntactic features.

The dependency inventory includes labels tailored to Modern Uyghur structures. Among the key labels are cop, used for explicit copula forms such as idi and iken; cop:zero, used for zero-copula predicative NP or AdjP structures; and three instrumental subtypes: instr:case=loc, instr:case=dat, and instr:case=post, corresponding respectively to instrumental marking by locative case suffix, dative case suffix, and explicit postposition such as bilen or arqiliq (Zuo et al., 29 Jul 2025).

MUDT also formalizes case relations by treating each case=X relation as a subtype of nmod, with notation such as case:loc, case:dat, case:abl, and case:poss (Zuo et al., 29 Jul 2025). This decision is consequential because it preserves distinctions that would otherwise collapse into more generic oblique or nominal-modifier categories. A plausible implication is that the treebank is intended to support downstream systems that depend on fine-grained predicate-argument and adnominal role distinctions.

3. Annotation principles and structural constraints

The framework is grounded in nine annotation principles. These are: No Isolated Nodes, Single Root Node, Single Head, Non-crossing Arcs, Directed Arcs, Head-first Annotation, Whole-Word Arcs, Complex Predicate Handling, and Semantics-driven Annotation (Zuo et al., 29 Jul 2025). Together, these principles define both the formal well-formedness of the dependency trees and the interpretive strategy used when syntax and semantics are not perfectly aligned.

The first seven principles specify core dependency-tree constraints. Every token appears in the tree; exactly one root is permitted per sentence; each non-root token has exactly one head; arcs may not cross; arcs are directed from head to dependent; the main predicate is identified as root before other attachments; and dependencies connect full tokens rather than sub-token pieces (Zuo et al., 29 Jul 2025). These constraints place MUDT within a projective, single-head dependency paradigm.

The final two principles are especially distinctive. For complex predicate handling, in converb+aux chains, the lexical verb is the head and auxiliaries attach with aux. For semantics-driven annotation, where syntax is ambiguous, lexical semantics determine the correct relation (Zuo et al., 29 Jul 2025). These principles clarify that MUDT does not simply encode surface adjacency or morphotactic prominence; it prioritizes lexical predication and semantic role transparency.

This combination of projectivity, lexical-head preference, and semantic disambiguation gives the framework a strong internal logic. It also explains why MUDT diverges sharply from analyses that privilege auxiliary-headed structures or more weakly typed oblique relations. The framework’s constraints are therefore not merely annotation hygiene; they are the mechanism through which Uyghur-specific grammatical distinctions are stabilized.

4. Corpus basis and annotation workflow

The baseline resource for the study is UD_Uyghur-UDT, described as containing 3,456 sentences and approximately 40,236 tokens, with 17 universal UD POS tags only and no multi-layer morphology (Zuo et al., 29 Jul 2025). MUDT’s pilot phase consists of the same test set re-annotated under the MUDT guidelines as test_mudt.conllu (Zuo et al., 29 Jul 2025). This direct reuse of the test set is methodologically important because it enables controlled cross-standard comparison.

Preprocessing retained sentence segmentation and tokenization consistent with UD_Uyghur-UDT, while adding multi-layer morphological tagging comprising surface form, lemma, segments, and FEATS (Zuo et al., 29 Jul 2025). Annotation tools included CoNLL-U format editors and in-house scripts for FEATS consistency (Zuo et al., 29 Jul 2025). The workflow therefore preserves comparability at the segmentation level while introducing substantially richer internal token annotation.

For quality control, the study reports a sanity check in which a UD-trained parser reproduces UD gold with 100% accuracy. Inter-annotator agreement was targeted via spot checks as future work. The standard parsing metrics named are UAS and LAS (Zuo et al., 29 Jul 2025). The 100% sanity-check result is not a general performance claim about Uyghur parsing; it is a controlled verification that the evaluation pipeline reproduces the baseline annotation standard exactly when gold and predicted representations are aligned.

This workflow suggests a deliberate separation between two questions: whether the parser can reproduce a known annotation scheme, and whether that scheme captures the intended linguistic analysis. MUDT is positioned as addressing the second question.

5. Cross-standard evaluation and systematic divergence from UD

The empirical validation uses a pre-trained Universal Dependencies parser, specifically the uyghur-udt-ud-2.15 model, in two evaluation scenarios: test_udt.conllu evaluated as UD gold versus UD predicted, and test_mudt.conllu evaluated as MUDT gold versus the same UD predicted output (Zuo et al., 29 Jul 2025). This setup isolates the effect of annotation mismatch while keeping the parser fixed.

The quantitative results are stark. On UDT Gold vs. UDT Predicted, the parser achieves 100.00% LAS F1, 100.00% MLAS, and 100.00% BLEX. On MUDT Gold vs. UDT Predicted, the corresponding scores are 52.10% LAS F1, 42.76% MLAS, and 50.00% BLEX (Zuo et al., 29 Jul 2025). The study interprets this as a 47.90% divergence, quantifying where UD analysis systematically fails on MUDT phenomena.

The reported failure modes are specific. First, for case relations, UD’s generic obl is contrasted with MUDT’s case:dat, case:loc, and case:abl, which preserve semantics. Second, for compound predicates, UD treats the auxiliary as root, whereas MUDT promotes the lexical verb to root and attaches the auxiliary with aux. Third, for fixed expressions, UD labels quote-clauses as advcl, while MUDT uses obj plus fixed to disambiguate roles. Fourth, for postpositional phrases, UD sometimes inverts the head from noun to postposition, whereas MUDT enforces the postposition as head with instr:post (Zuo et al., 29 Jul 2025).

These results constrain interpretation. They do not show that UD is unusable in general; rather, they show that when evaluated against MUDT’s target analysis, a UD parser systematically diverges on phenomena that MUDT treats as structurally and semantically consequential. The controversy is therefore not over parser optimization but over representational adequacy.

6. Canonical analyses, downstream implications, and replicability

The framework’s examples operationalize its analytical commitments. For zero copula, the sentence “U oqutqutfi.” (“He is a teacher.”) is annotated with oqutqutfi as the root and cop:zero(U → oqutqutfi) (Zuo et al., 29 Jul 2025). For a compound predicate, “Men hargiz saqlap turalmigidakman!” is analyzed with saqlap as root, advmod(hargiz → saqlap), nsubj(Men → saqlap), and aux(turalmigidakman → saqlap) (Zuo et al., 29 Jul 2025). For coordination, “Men alma, anar, nefpyt jedim.” is analyzed with conj(alma → anar), conj(anar → nefpyt), root jedim, nsubj(Men → jedim), and obj(nefpyt → jedim) (Zuo et al., 29 Jul 2025). For a postpositional phrase, “… daraxlar bilan qaplangan quruqluq.” is annotated with obj(daraxlar → bilan), instr:post(bilan → qaplangan), root qaplangan, and nsubj(quruqluq → qaplangan) (Zuo et al., 29 Jul 2025).

These examples clarify that MUDT’s notion of semantic transparency is implemented through head selection and label refinement. The framework does not merely rename dependencies; it reorients the dependency structure in places where lexical predication, case semantics, or postpositional governance would otherwise be obscured.

The study also states that while the UD parser scores only 52.10% LAS on MUDT, a parser trained on MUDT is expected to recover the missing 47.9% faithful analyses, boosting both UAS/LAS by approximately 15–25 points in preliminary experiments, identified as future work (Zuo et al., 29 Jul 2025). Because this is framed as expected performance in future work, it should be treated as a stated projection rather than a finalized benchmark.

The downstream benefits identified are more accurate predicate-argument structures for machine translation, better alignment of case roles, finer case distinctions for information extraction and semantic role labeling, and explicit discourse relations for discourse and sentiment analysis (Zuo et al., 29 Jul 2025). The paper further presents MUDT as a replicable model for other morphologically rich, low-resource languages, with four key steps: multi-layer morphological decomposition, language-specific fine-grained dependency relations, clear annotation principles, and cross-standard validation against UD or other baselines (Zuo et al., 29 Jul 2025).

Taken together, these claims position MUDT as both a Uyghur-specific treebank framework and a methodological template. This suggests that its broader significance lies not only in the resulting resource but in the argument that morphosyntactic annotation for low-resource, morphologically complex languages may require principled divergence from universal schemes when semantic fidelity is the priority.