forked from km1994/nlp_paper_study
-
Notifications
You must be signed in to change notification settings - Fork 0
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
2. 【关于 GECToR】 那些你不知道的事 3. 【关于 NLP比赛】 那些你不知道的事 添加内容
- Loading branch information
Showing
18 changed files
with
449 additions
and
4 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Empty file.
Binary file added
BIN
+38 KB
information_extraction/event_extraction/MLBiNet/img/微信截图_20210714095648.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added
BIN
+163 KB
information_extraction/event_extraction/MLBiNet/img/微信截图_20210714101357.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added
BIN
+17.3 KB
information_extraction/event_extraction/MLBiNet/img/微信截图_20210714101905.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added
BIN
+21.5 KB
information_extraction/event_extraction/MLBiNet/img/微信截图_20210714102148.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added
BIN
+9.22 KB
information_extraction/event_extraction/MLBiNet/img/微信截图_20210714102517.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added
BIN
+40.4 KB
information_extraction/event_extraction/MLBiNet/img/微信截图_20210714102921.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added
BIN
+30.5 KB
information_extraction/event_extraction/MLBiNet/img/微信截图_20210714103124.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added
BIN
+32.8 KB
information_extraction/event_extraction/MLBiNet/img/微信截图_20210715094254.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added
BIN
+31.3 KB
information_extraction/event_extraction/MLBiNet/img/微信截图_20210715094337.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added
BIN
+39.7 KB
information_extraction/event_extraction/MLBiNet/img/微信截图_20210715094543.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
151 changes: 151 additions & 0 deletions
151
information_extraction/event_extraction/MLBiNet/readme.md
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,151 @@ | ||
# 【关于 MLBiNet】那些你不知道的事 | ||
|
||
> 作者:杨夕 | ||
> | ||
> 论文:MLBiNet: A Cross-Sentence Collective Event Detection Network | ||
> | ||
> 会议: ACL2021 | ||
> | ||
> 论文下载地址:https://arxiv.org/pdf/2105.09458.pdf | ||
> | ||
> 论文代码:https://github.com/zjunlp/DocED | ||
> | ||
> 本文链接:https://github.com/km1994/nlp_paper_study | ||
> | ||
> 个人介绍:大佬们好,我叫杨夕,该项目主要是本人在研读顶会论文和复现经典论文过程中,所见、所思、所想、所闻,可能存在一些理解错误,希望大佬们多多指正。 | ||
> | ||
> 【注:手机阅读可能图片打不开!!!】 | ||
- [【关于 MLBiNet】那些你不知道的事](#关于-mlbinet那些你不知道的事) | ||
- [一、引言](#一引言) | ||
- [二、背景知识](#二背景知识) | ||
- [2.1 什么是 事件抽取?](#21-什么是-事件抽取) | ||
- [2.2 事件触发词检测任务面临的挑战 是什么?](#22-事件触发词检测任务面临的挑战-是什么) | ||
- [2.3 目前 事件抽取 存在问题?](#23-目前-事件抽取-存在问题) | ||
- [三、论文介绍](#三论文介绍) | ||
- [3.1 论文动机](#31-论文动机) | ||
- [3.2 论文方法](#32-论文方法) | ||
- [3.2.1 Semantic Encoder 语义编码器](#321-semantic-encoder-语义编码器) | ||
- [3.2.2 Forward Decoder 双向解码器](#322-forward-decoder-双向解码器) | ||
- [3.2.3 Information Aggregation 信息整合层](#323-information-aggregation-信息整合层) | ||
- [3.2.4 Multi-Layer Bidirectional Network 多层双向打标器](#324--multi-layer-bidirectional-network-多层双向打标器) | ||
- [3.2.5 Loss Function](#325-loss-function) | ||
- [四、实验结果](#四实验结果) | ||
- [参考](#参考) | ||
|
||
## 一、引言 | ||
|
||
We consider the problem of collectively detecting multiple events, particularly in cross-sentence settings. The key to dealing with the problem is to encode semantic information and model event inter-dependency at a document-level. In this paper, we reformulate it as a Seq2Seq task and propose a Multi-Layer Bidirectional Network (MLBiNet) to capturethe document-level as sociation of events and semantic information simultaneously. Specifically, a bidirectional decoder is firstly devised to model event inter-dependency with in a sentence when decoding the event tag vector sequence. Secondly, an information aggregation module is employed to aggregate sentence-level semantic and event tag information. Finally, we stack multiple bidirectional decoders and feed cross-sentence information, forming a multi-layer bidirectional tagging architectureto iteratively propagate information across sentences. | ||
|
||
- 动机:跨句事件抽取旨在研究如何同时识别篇章内多个事件 | ||
- 论文方法:论文将其重新表述为 **Seq2Seq 任务**,并提出了一个多层双向网络 (Multi-Layer Bidirectional Network,MLBiNet) 来 **融合跨句语义和关联事件信息,从而增强内各事件提及的判别** | ||
- 论文思路: 在解码事件标签向量序列时 | ||
- 首先,为建模句子内部事件关系,我们提出双向解码器用于同时捕捉前向和后向事件依赖; | ||
- 然后,利用信息聚合器汇总句子语义和事件提及信息; | ||
- 最后,通过迭代多个由双向解码器和信息聚合器构造的单元,并在每一层传递邻近句子的汇总信息,最终感知到整个文档的语义和事件提及信息。 | ||
|
||
## 二、背景知识 | ||
|
||
### 2.1 什么是 事件抽取? | ||
|
||
- 事件抽取组成: | ||
- 事件触发词检测(识别事件触发词,并明确所触发事件的类型) | ||
- 属性抽取(识别触发事件的属性,并标注各属性对应角色) | ||
|
||
> 注:“He died in hospital” 中 “died” 作为一个 Die 类型事件的触发词,该事件中,属性 “He” 的角色为 Person, “hospital” 的角色为 Place. | ||
### 2.2 事件触发词检测任务面临的挑战 是什么? | ||
|
||
1. 句子上下文表示及篇章级信息整合[1],[2]。**候选触发词类型的判定一般需要结合上下文信息**,包括关联实体信息(类型等)、其他候选触发词等。 | ||
|
||
> 例如,图 1 中句子 3 中的 “firing” 可能是开枪(触发 Attack 事件)或离职(触发 End_Position 事件),Attack 事件的确立需要融合句子2,4等的信息。 | ||
2. **句内和句间事件关联性建模**[1],[3]。句4包含事件触发词fight和death,ACE05数据集中超过40%触发词如此共现;类似句2、句3和句4中的连续关联事件同样普遍。 | ||
|
||
因此,建模事件之间依赖对于同时抽取句子、跨句多事件尤为重要。 | ||
|
||
![图 1 ](img/微信截图_20210714095648.png) | ||
|
||
### 2.3 目前 事件抽取 存在问题? | ||
|
||
现有方法主要专注于句子级事件抽取,忽略了存在于其他句子中的信息。 | ||
|
||
## 三、论文介绍 | ||
|
||
### 3.1 论文动机 | ||
|
||
1. 可将事件触发词检测任务视为一个**Seq2Seq任务**,对应基于RNN的encoder-decoder框架能有效处理该类问题,其中encoder建模丰富的上下文语义信息,decoder在解码过程中捕捉标签的依赖性。 | ||
1. source序列为文本篇章或句子; | ||
2. target序列是事件标签序列。 | ||
2. 对于当前句子,与之关联最密切的信息主要存在于邻近句子,相距较远的文本影响较小。 | ||
|
||
### 3.2 论文方法 | ||
|
||
![](img/微信截图_20210714101357.png) | ||
> 模型结构 | ||
#### 3.2.1 Semantic Encoder 语义编码器 | ||
|
||
- 结构:由BiLSTM和自注意力机制构成; | ||
|
||
![](img/微信截图_20210714101905.png) | ||
|
||
#### 3.2.2 Forward Decoder 双向解码器 | ||
|
||
- 结构:融合前向解码和后向解码,有助于捕捉双向事件依赖关系; | ||
|
||
![](img/微信截图_20210714102148.png) | ||
|
||
#### 3.2.3 Information Aggregation 信息整合层 | ||
|
||
- 结构:基于简单 LSTM 结构整合句子内部事件标签信息和语义信息 | ||
|
||
![](img/微信截图_20210714102517.png) | ||
|
||
#### 3.2.4 Multi-Layer Bidirectional Network 多层双向打标器 | ||
|
||
- 作用:多层双向打标器则逐层传递邻近句子信息,最终捕捉更大邻域范围内的语义和事件信息,进而实现跨句事件联合抽取。 | ||
- 结构主要约束包括: | ||
- (1)信息传递只发生在相邻句子间; | ||
- (2)当前句子中的所有token可见跨句信息是相同的; | ||
- (3)随着层数增加,较远距离的句子信息可被当前句子获取到; | ||
- (4)每层的双向打标器都由一个双向解码器和一个信息整合层构成 | ||
- 对于第 k 层事件标签向量信息计算方法为: | ||
|
||
![](img/微信截图_20210714102921.png) | ||
|
||
#### 3.2.5 Loss Function | ||
|
||
- negative log-likelihood loss function J(θ) | ||
|
||
![](img/微信截图_20210714103124.png) | ||
|
||
## 四、实验结果 | ||
|
||
在ACE05数据集上进行了试验,如下两个表所示,我们的方法在不同维度都能取得较好的效果。 | ||
|
||
- 结论: | ||
- 双向解码器有效,它在1层时较之于HBTNGMA更优; | ||
- 跨句信息整合有意义,多层网络下,我们的方法在单事件句子和多事件句子的抽取效果都得到提升。 | ||
|
||
![](img/微信截图_20210715094254.png) | ||
|
||
![](img/微信截图_20210715094337.png) | ||
|
||
模块剖析进一步了验证双向解码器和信息整合层的作用。 | ||
|
||
- 结论: | ||
- 双向解码器较之于单向方法显著更优; | ||
- 层数增加情况下,不同解码机制下的效果都能得到提升; | ||
- 不同信息整合机制也能引起一定表现变动。 | ||
|
||
![](img/微信截图_20210715094543.png) | ||
|
||
|
||
## 参考 | ||
|
||
1. Collective event detection via a hierarchical and bias tagging networks with gated multi-level attention mechanisms. EMNLP2018 | ||
2. Document embedding enhanced event detection with hierarchical and supervised attention. ACL2018 | ||
3. Jointly multiple events extraction via attention-based graph information aggregation. EMNLP2018 | ||
4. [ACL2021 | 探讨跨句事件联合抽取问题](https://mp.weixin.qq.com/s/Y3s8jvpKx-EoFOqa_ZCVfw) | ||
|
Oops, something went wrong.