Transition-based Dependency Parsing Using Recursive Neural Networks
Transition-based Dependency Parsing Using Recursive Neural Networks
复制标题
DOI：
--
复制
发表时间：
2013
影响因子：
7.2
通讯作者：
Pontus Stenetorp
中科院分区：
文献类型：
--
作者：
Pontus Stenetorp
在这项工作中，我们提出了一个通用的组合向量框架transitionbased依赖分析。使用基于转换的算法的能力允许将向量组合应用于只有依赖树库可用的大量语言，以及处理语言现象，例如对先前提出的方法造成问题的非投射性。我们引入了一个过渡有向无环图的概念，使我们能够应用递归神经网络与现有的基于过渡的算法进行解析。我们的框架捕捉短语之间的语义相关性类似于一个基于选民的对应文献，例如预测，“金融危机”，“现金紧缩”和“熊市”是语义相似。目前，基于我们的框架的解析器能够实现86.25%的未标记的附件分数为一个完善的依赖数据集，只使用单词表示作为输入，下降不到2%点，以前提出的可比的基于特征的模型。
In this work, we present a general compositional vector framework for transitionbased dependency parsing. The ability to use transition-based algorithms allows for the application of vector composition to a large set of languages where only dependency treebanks are available, as well as handling linguistic phenomena such as non-projectivities which pose problems for previously proposed methods. We introduce the concept of a Transition Directed Acyclic Graph that allows us to apply Recursive Neural Networks for parsing with existing transition-based algorithms. Our framework captures semantic relatedness between phrases similarly to a constituency-based counterpart from the literature, for example predicting that “a financial crisis”, “a cash crunch” and “a bear market” are semantically similar. Currently, a parser based on our framework is capable of achieving 86.25% in Unlabelled Attachment Score for a well-established dependency dataset using only word representations as input, falling less than 2% points short of a previously proposed comparable feature-based model.