Rethinking and Collaborative Learning Enhanced Cross-lingual Dependency Parsing

ACL ARR 2025 May Submission3000 Authors

Abstract: Large language models (LLMs) have shown strong syntax understanding capability in rich-source languages. However, their performances decline sharply when directly apply to low-resource languages. The key challenge is the data deviation and weak alignment across the source and target languages. To alleviate these issues, we propose a novel rethinking and collaborative learning approach for cross-lingual dependency parsing. On the one hand, we exploit a progressive thinking technique to guide LLMs to generate diverse and aligned synthetic data, thus making up for the data shift drawback. On the other hand, we introduce a collaborative learning strategy to further activate the alignment ability of both traditional cross-lingual models and LLMs by making full use of our synthetic data. Experiments on various benchmark datasets show that our proposed method outperform all strong baselines, leading to new state-of-the-art results on all language. Detailed comparison demonstrates that our synthetic data is extremely useful for enhancing the alignment between source and target languages. In-depth analysis reveals that both rethinking and collaborative learning can boost the cross-lingual parsing performance.

Paper Type: Long

Research Area: Syntax: Tagging, Chunking and Parsing

Research Area Keywords: mutil-lingual,LLM,dependency parsing,low-resource languages

Contribution Types: Approaches to low-resource settings, Data resources, Data analysis

Languages Studied: English,Chinese,Vietnamese,Tamil,Telugu,Maltese

Submission Number: 3000