Consistency-Constrained Joint Learning of Dependency, Constituency, and Grammatical Roles in Low-Resource Uzbek

Complementary syntactic annotations may support learning in low-resource languages, but compatibility penalties require evaluation beyond parameter sharing. We study dependency parsing, constituency parsing, grammatical-role labeling and chunking in Uzbek using a shared XLM-R encoder, structured predictors and empirical implication penalties during training. UzSyn contains 1,123 sentences and 4,955 tokens, partitioned into 898 training, 112 development and 113 test sentences. Comparisons used five seeds, independent models and unconstrained multi-task learning. With all training sentences, the constrained model reached 85.53% labeled attachment score and 76.55% constituency F1. Differences from unconstrained multi-task learning were +0.52 and +0.06 percentage points; paired 97.5% confidence intervals of [−1.71, 2.93] and [−2.08, 2.52] included zero. At the exploratory 50% final-fitting fraction, constituency and chunk F1 increased by 4.04 and 2.97 points, with descriptive 95% intervals excluding zero. Fitting budgets excluded prior annotation access during rule-template design. Coherence changes varied by rule family. Constrained fitting required 3.66 times the mean elapsed fitting time of unconstrained multi-task learning. The findings identify intermediate-budget gains for replication while leaving full-data superiority unresolved. Independent linguistic validation and external evaluation are needed before broader accuracy or annotation-saving claims.