Two views of linguistic structure

Context-free grammars (CFGs)

Phrase structure organizes words into nested constituents.

Dependency structure

Dependency structure shows which words depend on (modify, attach to, or are arguments of) which other words.

modify：修饰词，attach to：连接词

Why do we need sentence structure?

  • Humans communicate complex ideas by composing words together into bigger units to convey complex meanings.
  • Listeners need to work out what modifies [attaches to] what
  • A model needs to understand sentence structure in order to be able to interpret language correctly

Linguistic Ambiguities

  • Prepositional phrase attachment ambiguity 介词短语附着歧义
  • Coordination scope ambiguity 对等范围歧义
  • Adjectival/Adverbial Modifier Ambiguity 形容词修饰语歧义
  • Verb Phrase (VP) attachment ambiguity 动词短语依存歧义

Dependency paths identify semantic relations 依赖路径识别语义关系 help extract semantic interpretation.

Dependency Grammar and Dependency Structure

Tesnière had them point from head to dependent – we follow that convention.

We usually add a fake ROOT so every word is a dependent of precisely 1 other node.

箭头指向依赖者 dependent。

通常添加⼀个伪根指向整个句子的头部，这样每个单词都精确地依赖于另⼀个节点。

Dependency Conditioning Preferences

  • Bilexical affinities 两个单词间的密切关系
  • Dependency distance 依赖距离：主要是与相邻词
  • Intervening material 介于中间的物质：如果中间的词语是动词或者标点，则两边的词语不太可能有依存
  • Valency of heads：How many dependents on which side are usual for a head?

Methods of Dependency Parsing

Arc-standard transition-based parser 标准弧转移算法

基于转换的算法有一个栈和一个词队列，除此之外还有一个已经 parse 的依存关系。

该算法包含以下操作：

  • LEFT-ARC 栈顶和它下面的词构成依存关系，并且中心词是栈顶元素，把这两个词从栈中弹出，把这个依存关系加入到已parse的数据结构里，最后把中心词再加到栈中
  • RIGHT-ARC 栈顶和它下面的词构成依存关系，中心词是下面的元素，把这两个词从栈中弹出，把这个依存关系加入到已parse的数据结构里，最后把中心词再加到栈中
  • SHIFT 把队列中的一个词加入到栈顶

这个算法很简单，初始状态栈里只有一个root，而队列里是所有的词，然后循环直到状态是结束状态(栈中只有root，队列也为空)。

例子：

Other Algorithms

  • Greedy transition-based parsing
  • MaltParser
  • Conventional Feature Representation

Evaluation of Dependency Parsing

  • UAS (unlabeled attachment score) 指 无标记依存正确率 ，不考虑依存关系的类型，只考虑依存关系的对应单词是否正确，如父节点被正确识别的准确率；
  • LAS (labeled attachment score) 指有标记依存正确率，同时考虑依存关系的对应单词以及依存关系类型，即词的父节点，以及与父节点的句法关系都被正确识别的概率。

Neural dependency parsing

Just one word win.