Learning to Search for Dependencies
Kai-Wei Chang1, He He2, Hal Daumé III2, John Langford3 1 University of Illinois Urbana-Champaign, IL 2 University of Maryland, College Park, MD 3 John Langford Microsoft Research, New York, NY
Abstract
We demonstrate that a dependency parser can be built using a credit assignment compiler which removes the burden of worrying about low-level machine learning details from the parser implementation. The result is a simple parser which robustly applies to many languages that provides similar statistical and computational performance with best-to-date transition-based parsing approaches, while avoiding various downsides including randomization, extra feature requirements, and custom learning algorithms.
中文速览
用"学习搜索编译器"（learning to search compiler）来自动处理依存句法分析（dependency parsing）中棘手的信用分配问题，可以把解析器的实现大幅简化——开发者只需写好解码器和动态oracle，剩下的机器学习细节全交给编译器搞定。作者基于这一思路，用约300行C++代码构建了一个弧混合转移式依存解析器，在英语Penn Treebank和CoNLL-X十个语言数据集上的表现与当前最优的转移式解析方法相当甚至更好（十语言平均标注准确率81.7，对比基线的80.3）。这套方案不需要随机化技巧、额外特征工程或定制学习算法，还能通过命令行参数直接切换神经网络等高级学习器。它的意义在于为依存句法分析提供了一个既简洁又强健的基线，同时也证明"编译器驱动"的结构化预测思路完全可以推广到其他复杂NLP任务。
原文 arXiv:1503.05615;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1503.05615v2