Research map / Label-efficient and robust learning / Recognition and long-tail
Dynamic Syntactic Feature Filtering and Injecting Networks for Cross-lingual Dependency Parsing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a shared-private cross-lingual dependency parser with a language-specific filtering network (LSFN) on private BiLSTMs and a language-invariant injecting network (LIIN) on the shared encoder, plus KAN scoring, evaluated on seven low-resource UD languages.
Problem
Cross-lingual dependency parsing for low-resource languages suffers from syntactic structure distributional bias from source languages and difficulty capturing deep cross-language commonalities.
Approach
Filter helpful source-language syntactic features in private encoders via learnable layer and element weights that mimic source representations, and inject self-attention plus Mamba layers into the shared BiLSTM encoder to capture long-range language-invariant syntax; KAN replaces MLP in the decoder.
Key results
- Average gain of 1.84 UAS and 3.43 LAS over the shared-private model (ShaPri) across seven languages; average LAS 74.00 vs 70.57 for ShaPri.
- On Vietnamese dev, the full model reaches 76.53 UAS / 61.65 LAS vs 75.04 / 59.52 for the shared-private baseline with KAN.
- Adds about 9.4M parameters over the baseline and about 14 seconds of training time on Vietnamese per the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed dynamic filtering and injecting networks instead of ShaPri | Cross-lingual dependency parser · Seven-language test average | UAS: +1.84 UAS | Helped |
| Use the proposed dynamic filtering and injecting networks instead of ShaPri | Cross-lingual dependency parser · Seven-language test average | LAS: +3.43 LAS | Helped |
| Add LSFN to the full method (ablation) | Proposed parser · Vietnamese dev | UAS: 75.58 -> 76.53 | Helped |
| Add LSFN to the full method (ablation) | Proposed parser · Vietnamese dev | LAS: 60.90 -> 61.65 | Helped |
| Add LIIN to the full method (ablation) | Proposed parser · Vietnamese dev | UAS: 75.81 -> 76.53 | Helped |
| Add LIIN to the full method (ablation) | Proposed parser · Vietnamese dev | LAS: 60.30 -> 61.65 | Helped |
| Add both LSFN and LIIN to the full method (ablation) | Proposed parser · Vietnamese dev | UAS: 75.04 -> 76.53 | Helped |
| Add both LSFN and LIIN to the full method (ablation) | Proposed parser · Vietnamese dev | LAS: 59.52 -> 61.65 | Helped |
| Use KAN instead of MLP (ablation) | FulSha · Vietnamese dev | UAS: 75.04 -> 74.83 | Hurt |
| Use KAN instead of MLP (ablation) | FulSha · Vietnamese dev | LAS: 57.63 -> 57.89 | Helped |
| Use KAN instead of MLP (ablation) | ShaPri · Vietnamese dev | UAS: 75.48 -> 75.04 | Hurt |
| Use KAN instead of MLP (ablation) | ShaPri · Vietnamese dev | LAS: 59.24 -> 59.52 | Helped |
Surprising findings
- Removing LSFN or LIIN individually drops LAS by a similar margin, and removing both causes the largest drop, with the authors describing the two components as complementary.
Limitations
- Gains are reported on a single source language per target and rely on small treebanks for target languages.
- Improvements over ShaPri are modest for some languages in the ablation, and the approach depends on hyperparameter choices such as loss weight alpha fixed at 1.
Experimental setup
Datasets: UD English EWT UD Chinese GSDSimp UD Vietnamese VTB UD Wolof WTB UD Coptic Scriptorium UD Maltese MUDT UD Tamil TTB UD Uyghur UDT TUD Thai
Benchmarks: UD Vietnamese VTB UD Wolof WTB UD Coptic Scriptorium UD Maltese MUDT UD Tamil TTB UD Uyghur UDT TUD Thai
Metrics: UAS LAS
Baselines: UDPipe UDify MBERT ESR FulSha LanEmb ShaPri
Builds on: shared-private model BiAffine parser Mamba Kolmogorov-Arnold Network XLM-RoBERTa
Closest papers by meaning
- Multilingual Pre-training with Universal Dependency Learning (NeurIPS 2021 · similarity 0.88)
- A Representation Learning Framework for Multi-Source Transfer Parsing (AAAI 2016 · similarity 0.87)
- Dependency Parsing is More Parameter-Efficient with Normalization (NeurIPS 2025 · similarity 0.86)
- Implicit Word Reordering with Knowledge Distillation for Cross-Lingual Dependency Parsing (AAAI 2025 · similarity 0.86)
- Enhancing Multilingual LLM Pretraining with Model-Based Data Selection (NeurIPS 2025 · similarity 0.85)
- Breaking Physical and Linguistic Borders: Multilingual Federated Prompt Tuning for Low-Resource Languages (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.