Graph-based approach is one of the most successful approaches to dependency parsing.It is attractive for capability of offering global inference over space of all possible trees,and thus guarantees to find the best-scored trees given a tree scoring model.Traditional graph-based dependency parsing models usually adopt linear feature-based scoring models,which heavily rely on time-consuming feature engineering.The huge number of features they involves also dramatically slows down the parsing speed.Typical graph-based models factor the dependency tree into subgraphs,which limits the scope of feature extraction to the subgraph and inhibits the performance of recovering long distance dependencies.Recent work introduces deep learning models into graph-based dependency parsing models and seems to partially solve or alleviate the problems.In the paper Ⅰ survey some of this work and present the advances they have achieved.%图解码依存分析方法是一种重要的依存分析方法,优点是解码具有全局最优的特点,能够找到模型意义下的全局最佳依存树.传统图解码依存分析模型大多采用基于特征的线性评分模型,常常需要选取大量的人工特征,这一方面耗时费力,加剧了模型过拟合的风险,另一方面也显著降低了系统的运行效率.同时由于采用子图分解策略,传统图解码分析中的特征提取严重受到子图规模的限制,无法提取具有全局意义的分析特征.深度图解码依存分析研究部分解决了这些问题,本文概要介绍了近年来几个代表性的深度图解码依存分析研究工作,总结了国内外在深度图解码依存分析方面的现状和进展.
展开▼