UniDive 2025 Morpho-Syntactic Parsing

• The paper introduces a joint multitask neural architecture using XLM-RoBERTa and character-level embeddings to address morpho-syntactic parsing for typologically diverse languages.
• It details a unified pipeline that performs gold-standard tokenization, content word identification, dependency parsing, and multi-label morphological feature prediction with precise evaluation metrics.
• Results demonstrate robust accuracy in most languages while highlighting challenges like segmentation ambiguity in Turkish and Hebrew, informing future improvements.

UniDive 2025 Morpho-Syntactic Parsing is a benchmark and shared task focused on the joint prediction of morphological and syntactic structure over typologically diverse languages, following an enhanced Universal Dependencies (UD) scheme. It targets the simultaneous assignment of syntactic dependencies and rich, multi-label morphological features, including content–function word discrimination and morpho-syntactic feature assignment, all over gold-standard tokenization. The 2025 edition stimulated novel architectures, compared pipelines, joint, and hybrid approaches, and provided fine-grained analyses, particularly for morphologically rich languages and challenging grammatical phenomena (Inostroza et al., 19 Aug 2025).

1. Task Definition and Shared Task Objectives

UniDive 2025 requires a parsing system that, given raw text input in any of nine target languages, performs the following subtasks in a unified and consistent way:

• Tokenization: Output sequences must conform to gold-standard tokenization. Token boundaries often correspond to morpheme segmentation in morphologically rich languages.
• Content Word Identification (CWI): Each token must be labeled as content or function word, a necessary step for focusing downstream parsing and morphological analyses on informative units.
• Dependency Parsing: Output a projective dependency tree over content words, with appropriate labeled syntactic relations.
• Morphosyntactic Feature Prediction: Assign a complete multi-label set of morphological features (e.g. Case, Number, Gender, Tense) for each content word.

The evaluation spans nine languages (Czech, English, Hebrew, Italian, Polish, Portuguese, Serbian, Swedish, Turkish), emphasizing typological diversity and varying degrees of morphological complexity. Output is scored using three principal metrics: MSLAS (Morpho-Syntactic Labeled Attachment Score), LAS (Labeled Attachment Score), and Feats F1, all computed using the shared task scoring protocol (Inostroza et al., 19 Aug 2025).

2. Joint Multitask Model Architecture

The leading UniDive 2025 system applies a joint multitask neural architecture with a multilingual XLM-RoBERTa encoder, character-level augmentation (via Flair), and three task-specific decoders. The workflow proceeds as follows:

• Input text is tokenized by a Stanza model (best-matched per language; e.g., HTB for Hebrew), producing token sequences.
• Tokens are embedded using XLM-RoBERTa_large, plus concatenated character embeddings to improve handling of rare morphological forms.
• All token representations pass through a shared intermediate layer comprising a linear transform, ReLU activation, and dropout.
• Three parallel decoders operate on the shared representations:
  • CWI Decoder: BiLSTM (256 units per direction) → locked dropout (p=0.5) → linear classification (content vs. function). Loss: class-weighted cross-entropy with an additional 5% token-level embedding dropout.
  • Dependency Parsing Decoder: Only content tokens are fed to an MLP_arc (256 units) + MLP_rel (128 units), using ReLU and LayerNorm. Parsing uses a Dozat & Manning biaffine attention mechanism and structured inference via a projective tree CRF (TorchStruct). Loss: negative log-likelihood over projective trees.
  • Morphosyntactic Feature Predictor: Linear layer with sigmoid activation for multi-label prediction across all possible (feature=value) pairs; threshold at 0.5. Function words bypass this component (assigned “_”). Loss: summed binary cross-entropy.

Training employs a weighted sum of decoder losses:

$$L_\text{total} = w_\text{parser} \cdot L_\text{parser} + w_\text{morph} \cdot L_\text{morph} + w_\text{CWI} \cdot L_\text{CWI}$$

Typical weights are (2.0:1.5:1.0) for parser:morph:CWI, with higher morph/CWI weights for Turkish and Swedish.

Regularization involves dropout at multiple stages. Optimizer: AdamW (weight decay 0.01, initial lr=2×10⁻⁵, batch size=16). Training on NVIDIA A100 completes in 1–5 hours per language; selection is via best seed after three initial developmental seeds (Inostroza et al., 19 Aug 2025).

3. Preprocessing, Tokenization, and Postprocessing

Tokenization is performed with pretrained Stanza models, chosen per language by full-pipeline evaluation for best dev-set MSLAS. Gold tokenization is strictly matched at both training and test time. For ambiguous cases, such as low-confidence content/function predictions, postprocessing corrects CWI label inversion when a token is flanked by the opposite label and enforces at least one content word in short sentences.

This pipeline approach yields robust token-level alignment that is critical for system performance, especially for languages exhibiting segmentation ambiguity and high degree of fused morphology (e.g., Hebrew, Turkish). A plausible implication is that segmentation–content identification interdependence is a principal challenge for morpho-syntactic parsing in such languages (Inostroza et al., 19 Aug 2025).

4. Evaluation Protocol and Empirical Results

Systems are evaluated by three principal metrics:

• Labeled Attachment Score (LAS):

$$\mathrm{LAS} = \frac{|\{t : \text{head}_\text{pred}(t) = \text{head}_\text{gold}(t) \land \text{rel}_\text{pred}(t) = \text{rel}_\text{gold}(t)\}|}{N}$$

• Feats F1: Micro-averaged F1 over (feature=value) matches.
• MSLAS (Morpho-Syntactic LAS): Feats F1 restricted to tokens with correct head and relation.

The official test set results are summarized below:

Language	MSLAS	LAS	Feats F1
Czech	87.1	88.0	95.2
English	83.8	85.1	94.9
Hebrew	68.7	71.4	83.4
Italian	73.0	73.7	84.7
Polish	75.0	76.5	86.2
Portuguese	88.9	89.5	94.8
Serbian	86.6	88.3	95.6
Swedish	86.6	87.7	95.7
Turkish	58.7	60.9	82.1
Average	78.7	80.1	90.3

Ablation studies show that gold tokenization provides MSLAS gains up to +9.3 (Hebrew), and gold content/function labeling yields a further +1 point on average. The results demonstrate robust generalization across typologically diverse languages, but note lower performance in Turkish and Hebrew, where morphological ambiguity and segmentation errors remain limiting (Inostroza et al., 19 Aug 2025).

5. Comparative Systems and Methodological Innovations

Several research directions in joint or hybrid morpho-syntactic parsing were prominent in the UniDive 2025 context:

• Joint Lattice-based Architectures: Directly encode all segmentation and tagging ambiguities as a lattice and solve segmentation and parsing jointly via constrained MST inference. This approach has demonstrated state-of-the-art performance on Hebrew (e.g., SEG F1=97.71, POS F1=94.41, DEP F1=85.45), but is computationally intensive due to lattice size (Levi et al., 2024). Adapting these systems for UniDive requires a language-agnostic morphological analyzer and appropriate LLM encoder (e.g., mBERT, AlephBERT).
• Flipped Pipelines: Use parallel, independent expert classifiers for segmentation, morphology, lemmatization, and dependency parsing over space-delimited tokens. The synthesis step combines predictions into a fully segmented UD structure. This method dramatically reduces computation time (e.g., 0.0019 s/sentence vs. 0.257 s/sentence for joint lattice models) and achieves near SOTA accuracy (e.g., DictaBERT-Parse-large on Hebrew: UAS 89.50, LAS 85.40), but decouples token-level decisions from syntax (Shmidman et al., 2024).
• Resource-Integrated Approaches: Enrich token representations by fusing multiple sources of morpho-syntactic knowledge. For instance, UniDive for Korean integrates UD and UniMorph features, yielding +14.09 LAS improvement with explicit morpho-syntactic feature input (Park et al., 26 Mar 2025). Targeted ablation confirms that functional morpheme features are the main drivers of these gains. Explicit Case features reduce subject/object confusion and syntactic attachment errors.

A summary comparison appears below:

System Type	Notable Feature	Parsing Accuracy (Hebrew/Turkish)	Computational Profile
Joint Multitask	Shared encoder + task-specific decoders	71.4 (Hebrew LAS), 60.9 (Turkish LAS)	Moderate (1–5 hr/language GPU) (Inostroza et al., 19 Aug 2025)
Lattice-Based	Joint segmentation & parsing	DEP F1: 85.45–86.55 (Hebrew)	High (lattice combinatorial) (Levi et al., 2024)
Flipped Pipeline	Independent expert modules	LAS: 84.7–85.4 (Hebrew)	Fastest, highly parallel (Shmidman et al., 2024)
Resource-Integrated	Fused morphosyntactic features	LAS: +14 using MS features (Korean)	Task-dependent (Park et al., 26 Mar 2025)

6. Error Analysis and Linguistic Phenomena

Fine-grained analyses reveal consistent areas for future focus:

• Nominal Morphology: The most common errors involve confusion between core grammatical cases, particularly Nominative vs. Accusative (“Nom-Acc”). For instance, the model might predict Case=Nom for an Accusative-marked direct object, especially in languages with syncretism. Gender and Number are predicted almost perfectly in some languages (e.g. >99% in Czech), but syncretic forms sporadically yield invalid feature sets.
• Dependency Relations: Boundaries between certain relations, notably nmod (nominal modifier) and obl (oblique), account for over 40% of top relation confusions. Other challenges include obj/nsubj and nmod/flat confusion, especially for ambiguous or elliptical child language data.
• Attachment Distance and Direction: The model tends to predict longer dependency attachment distances in errorful parses and often preserves head directionality, indicating robust but not completely precise syntactic priors.

Error analysis validates that gold tokenization and content labeling directly underpin morpho-syntactic parsing performance, implying that advances in unsupervised or semi-supervised tokenization for MRLs will have disproportionate impact on parsing accuracy (Inostroza et al., 19 Aug 2025, Park et al., 26 Mar 2025).

7. Future Directions and Broader Impact

UniDive 2025 Morpho-Syntactic Parsing supports continued methodological innovation for both broad-coverage parsers and morphologically rich languages. Persistent challenges such as case confusion and relation boundary fuzziness motivate research into case-specific loss weighting and constrained decoding strategies.

Cross-lingual transfer, explicit feature integration, and robust tokenization are promising avenues, especially for low-resource and typologically diverse languages. Integration of automated UD morphosyntactic pipelines (e.g., for developmental speech in CHILDES (Liu et al., 2024)) demonstrates applicability beyond the core NLP research community, supporting linguistic, educational, and clinical analysis at scale.

Publicly released code and models catalyze reproducibility and extension, positioning UniDive as a reference for future morpho-syntactic parsing research (Inostroza et al., 19 Aug 2025).