动作n-gram模型提升Shift-Reduce成分解析精度

71 浏览量更新于2024-07-15 收藏 478KB PDF 举报

本文主要探讨了如何通过"Action N-Gram模型"来提升Shift-Reduce类型成分分析的性能。Shift-Reduce算法在自然语言处理中的成分分析（Constituent Parsing）是一种常见的方法，它依赖于大量的二元指示特征与区分性模型来理解句子结构。然而，传统的Shift-Reduce解析器往往面临上下文理解能力有限的问题。 Action N-Gram模型的提出旨在解决这一问题。该模型利用动作序列（Actions）作为辅助信息，帮助解析过程中对歧义进行更精准的判断。在训练阶段，模型采用n-gram估计方法，这种方法对给定历史规格（specification history）下动作概率进行平滑最大似然估计，从而提供更为准确的概率分布。与现有最先进的解析框架结合时，Action N-Gram模型展示了显著的优势。作者通过在两种不同语言的三个数据集上进行实验，证实了该模型能够提高解析的准确性。实验结果表明，通过动作n-gram模型的融入，解析系统的性能得到了实质性的提升，这不仅对于英语等语言，对于其他语言处理任务也有着广泛的应用潜力。具体来说，本文的工作属于人工智能领域中的自然语言处理子领域——句法分析，特别是在处理复杂语法结构和解决歧义问题方面。关键词包括"Shift-reduce"、动作序列、n-gram估计以及语言模型。通过这些技术改进，研究人员可以构建出更加智能和高效的句法解析器，为后续的文本分析、机器翻译、问答系统等应用奠定坚实基础。

13:4 H. Zhou et al.

accuracy than other parsers, which indicates that the nonlocal features are very helpful

to syntax parsing.

2.2. Using Action for Parsing Disambiguation

Briscoe and Carrol [1993] described work toward the construction of a probabilistic

parsing system for natural language, based on the LR parsing technique. They proposed

to associate probabilities with transitions in the automaton in a generalized LR parsing

framework [Tomita 1987]. They combined parsing action with the current parsing

state, lookahead item, and resultant nonterminal. The ﬁnal probabilities of combined

actions are derived from the set of parse histories resulting from the training phase,

by counting the frequencies of combined actions and converting these to probabilities.

There are many generalized LR parsers that assign probabilities to actions [Lavie

1996; Kentaro et al. 1998]. Compared to associating the probabilities to the rules of

the grammar, the methods allow the probabilistic parser to distinguish situations in

which identical rules reapply in different ways across different derivations or apply

with differing probabilities in different contexts.

The action n-gram model and statistical generalized LR parsers both assign prob-

abilities to actions for parsing disambiguation. The statistical generalized LR parser

assigns probabilities to the next actions by counting their co-occurrence frequencies

with the current state, lookahead item, and resultant state. In contrast, the action n-

gram model assigns probabilities to the next actions by counting the previous n actions

directly. In generalized LR parsing, the action is combined with the lookahead item and

resultant nonterminal in the LR parsing table. But the action of action n-gram models

is based on head word, head pos-tag, and constituent label information in the parsing

stack. The statistical generalized LR parser focuses on exploiting the state transitions

in parsing. In contrast, the action n-gram model focuses on the action s equence and

syntax tree structures formed by the action sequence.

The action n-gram model is built upon a data-driven shift-reduce parsing framework

with an n-gram estimation method rather than on a generalized LR parsing framework.

Compared to using the action history only, we propose to incorporate the action n-gram

model into a discriminative parsing model to enhance the parsing performance.

3. SHIFT-REDUCE PARSING

Typical shift-reduce parsers parse a sentence by performing a sequence of shift-reduce

actions. The action to be performed is determined by a statistical classiﬁer, and the

parsing result is obtained by searching greedily from left to right of the sentence

[Sagae and Lavie 2005]. Zhang and Clark [2009] applied global discriminative training

and beam-search to obtain higher accuracies. Zhu et al. [2013] added a new action to ﬁll

the gap among action sequences with different lengths caused by unary reduce actions.

In this section, we brieﬂy review the shift-reduce constituent parsing framework.

3.1. Actions of Shift-Reduce Parsing

A shift-reduce parser parses a sentence from left to right and generates the whole parse

tree by performing a sequence of actions. The parser starts with an initial state and

makes transitions from one to another by performing actions. Every state consists of

a queue Q{q

, q

,...} and a stack S{...,s

, s

}, where Q contains the word and

POS-tag pairs to be processed and S contains partially parsed subtrees. In each step,

one of the following actions is applied:

—SHIFT (S): push the ﬁrst word and POS-tag pair of the queue onto the stack as the

top node s

ACM Trans. Asian Low-Resour. Lang. Inf. Process., Vol. 15, No. 3, Article 13, Publication date: February 2016.

剩余16页未读，继续阅读

weixin_38517892

粉丝: 3
资源: 950

动作n-gram模型提升Shift-Reduce成分解析精度

N-gram语言模型

Action N-Gram模型提升Shift-Reduce成分分析

n-gram-tree:用Java编写的n-gram模型

n-gram:使用字符 N-gram 的电影评论语义分析

n-gram模型 分析汉语 python

使用jieba搭建n-gram模型python

使用py写一N-gram模型

n-gram模型训练代码

使用N-Gram模型基于特征扩展的短文本分类

nlp数据包 用于分词，n-gram模型，情感分析等

最新资源

n-gram模型分析汉语 python

nlp数据包用于分词，n-gram模型，情感分析等