论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension

文章目录

*
1 简介

+ 1.1 创新
2 任务定义
3 方法
4 模型
5 实验

1 简介

论文题目:Reading the Manual-Event Extraction as Definition Comprehension
论文来源:EMNLP 2020
论文链接:https://aclanthology.org/2020.spnlp-1.9.pdf

1.1 创新

  • 提出一个新的事件抽取方法,通过使用bleached statements来利用注释指南,同时使用多-span选择模型来进行事件抽取。
    论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension

; 2 任务定义

一个bleached statement包含状态token S = ( s 1 , s 2 , . . . , s n ) S=(s_1,s_2,…,s_n)S =(s 1 ​,s 2 ​,…,s n ​),占位符字典R = { ( r k : I k ) } k = 1 , . . . , K R={(r_k:I_k)}_{k=1,…,K}R ={(r k ​:I k ​)}k =1 ,…,K ​(分别表示论元角色和索引),一个例子如下:

论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension
论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension
将触发词定义为一种特殊的论元,因此,任务定义为给定一个bleached statement S、占位符词典R和文本tokenT,返回一个字典R ^ \widehat{R}R(包含触发词和论元,如下图右下角)
论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension

3 方法

给定带有多个占位符的bleached statement,增量式的填充每个占位符(从左到右逐步替换占位符,在训练中使用gold论元进行替换),公式为A ← G E T A R G S ( S , I , T ) A{\leftarrow}GETARGS(S,I,T)A ←G E T A R G S (S ,I ,T ),如果返回为空,则不替换占位符,如果返回为多个论元,用and连接,算法过程如下图:

论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension
完整的事件抽取算法如下,首先进行触发词检测,然后进行论元抽取。
[En]

The complete event extraction algorithm is as follows: first, trigger words are detected, and then arguments are extracted.

论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension

; 4 模型

首先使用BERT对bleached statement和文本编码,公式为BERT([CLS,s_1,…,s_n,SEP,t_1,…,t_m,SEP]),然后对文本进行BIO标注,公式如下:

论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension

将触发词当作特殊的论元,使用论元选择模块进行触发词识别,触发词不属于任何论元。在事件检测中,通过设置样本中未出现的事件类型,进行负采样。

[En]

The trigger word is regarded as a special argument and the argument selection module is used to identify the trigger word. The trigger word does not belong to any argument. In event detection, negative sampling is carried out by setting the type of event that does not appear in the sample.

同时使用SQuAD 2.0进行预训练,问题设置主要为wh-question phrases,使用AllenNLP toolkit抽取出wh-phrases。

5 实验

实验数据集为ACE 2005,实验结果如下图(使用部分数据为仅使用newswire的58个文档进行训练):

论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension
在FrameNet的Few-shot实验结果如下图:
论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension
在FrameNet的Zero-shot实验结果如下图:
论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension

Original: https://blog.csdn.net/o11oo11o/article/details/123609900
Author: hlee-top
Title: 论文笔记 EMNLP 2020|Reading the Manual-Event Extraction as Definition Comprehension

原创文章受到原创版权保护。转载请注明出处:https://www.johngo689.com/79287/

转载文章受原作者版权保护。转载请注明原作者出处!

(0)

大家都在看

最近整理资源【免费获取】:   👉 程序员最新必读书单  | 👏 互联网各方向面试题下载 | ✌️计算机核心资源汇总