Branching Out: Exploration of Chinese Dependency Parsing with Fine-tuned Large Language Models

In this paper, we investigate the effectiveness of large language models (LLMs) for Chinese dependency parsing through fine-tuning. We explore how different dependency representations impact parsing performance when fine-tuning the Chinese Llama-3 model. Our results demonstrate that while the Stanford typed dependency tuple representation yields the highest number of valid dependency trees, converting dependency structure into a lexical centered tree produces parses of significantly higher quality despite generating fewer valid structures. The results further show that fine-tuning enhances LLMs’ capability to handle longer dependencies to some extent, though challenges remain. Additionally, we evaluate the effectiveness of DeepSeek in correcting LLM-generated dependency structures, finding that it is effective for fixing index errors and cyclicity issues but still suffers from tokenization mismatches. Our analysis across dependency distances and relations reveals that fine-tuned LLMs outperform traditional parsers in specific syntactic structures while struggling with others. These findings contribute to the research on leveraging LLMs for syntactic analysis tasks.