跨资源区间的依存句法分析：评估高资源与低资源语言上的架构性能
文章背景与核心概要
本文探讨了依存句法分析模型在不同数据资源规模下的性能表现，重点对比了传统神经架构（如双仿射LSTM和栈指针网络）与预训练Transformer模型（AfroXLMR-large和RemBERT）。通过对12种语系各异的语言（着重关注资源匮乏的非洲语言）进行评估，研究表明，尽管Transformer在高资源场景下表现卓越，但在低资源环境下，双仿射LSTM的表现 consistently 优于它们。
此外，研究发现形态复杂度（通过MATTR衡量）是当控制语料库大小时影响Transformer性能的关键次要因素。这些发现为低资源语言的句法工具开发提供了重要参考，表明在积累到足以发挥预训练Transformer表征能力的标注数据之前，双仿射LSTM可能是更合适的选择。
Summary / 摘要
This paper investigates the performance of dependency parsing models across different data resource scales, specifically comparing traditional neural architectures (like the Biaffine LSTM and Stack-Pointer Network) with pre-trained transformer models (AfroXLMR-large and RemBERT). Evaluating twelve typologically diverse languages with an emphasis on under-resourced African languages, the study demonstrates that while transformers excel in high-resource scenarios, the Biaffine LSTM consistently outperforms them in low-resource regimes. Furthermore, morphological complexity (measured via MATTR) is identified as a key secondary factor influencing transformer performance when corpus size is controlled.
本文探讨了依存句法分析模型在不同数据资源规模下的性能表现，具体对比了传统神经架构（如双仿射LSTM和栈指针网络）与预训练Transformer模型（AfroXLMR-large和RemBERT）。通过对十二种语系各异的语言（重点关注资源匮乏的非洲语言）进行评估，研究表明，尽管Transformer模型在高资源场景下表现出色，但在低资源环境下，双仿射LSTM的表现始终优于Transformer。此外，在控制语料库大小时，形态复杂度（通过MATTR测量）被确定为影响Transformer性能的关键次要因素。
Metadata / 元数据
- arXiv ID: arXiv:2605.02608 [cs.CL]
- Subjects: Computation and Language (
cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)- Authors: Kevin Guan, Happy Buzaaba, Christiane Fellbaum
- Submission Dates: Submitted on 4 May 2026; Last revised 7 Aug 2026 (v2)
- DOI: 10.48550/arXiv.2605.02608
- arXiv ID: arXiv:2605.02608 [cs.CL]
- 研究主题: 计算与语言 (
cs.CL)；人工智能 (cs.AI)；机器学习 (cs.LG) - 作者: Kevin Guan, Happy Buzaaba, Christiane Fellbaum
- 提交日期: 2026年5月4日提交；最近修订于2026年8月7日（v2版）
- DOI: 10.48550/arXiv.2605.02608
Abstract / 摘要
Transformer-based models achieve state-of-the-art dependency parsing for high-resource languages, yet their advantage over simpler architectures in low-resource settings remains poorly understood. We evaluate four parsers—the Biaffine LSTM, Stack-Pointer Network, AfroXLMR-large, and RemBERT—across twelve typologically diverse languages, with a focus on low-resource African languages.
基于Transformer的模型在高资源语言的依存句法分析中达到了先进水平，然而在低资源环境下，它们相比更简单架构的优势仍未被充分理解。我们评估了四种解析器——双仿射LSTM、栈指针网络、AfroXLMR-large和RemBERT——在十二种语系各异的语言上的表现，重点关注低资源的非洲语言。
We find that the Biaffine LSTM consistently outperforms transformer models in low-resource regimes, with transformers recovering their advantage as training data increases. The crossover falls within a resource range typical of treebanks for under-resourced languages. Morphological complexity (measured via MATTR) emerges as a significant secondary predictor of transformers' relative disadvantage after controlling for corpus size. These results indicate that the Biaffine LSTM may be better suited for syntactic tool development in low-resource regimes until sufficient annotated data is available to leverage the representational capacity of pre-trained transformers.
我们发现，在低资源环境下，双仿射LSTM的表现始终优于Transformer模型，而随着训练数据的增加，Transformer重新获得了优势。这一性能交界点恰好落在了资源匮乏语言树库的典型资源范围内。在控制语料库大小后，形态复杂度（通过MATTR测量）成为预测Transformer相对劣势的一个显著次要因素。这些结果表明，在获得足够标注数据以利用预训练Transformer的表征能力之前，双仿射LSTM可能更适合低资源环境下的句法工具开发。
