Emotion-Cause Pair Extraction: A

2021-12-13  本文已影响0人  一个迷人的昵称

ECPE任务

区别:

微信图片_20211213111138.png
即`句子4`中的**anger**的情绪是由于`句子1`和`句子2`
image.png image.png

对比IndependentInter-ECInter-CE的结果,可以看到 Inter-~ 都会有所提升

the improvements of Inter-EC on the cause extraction task are much more than the improvement of InterCE on the emotion extraction task

it is because cause extraction is more difficult than emotion extraction, hence there is more room for extra improvement.

image.png

By comparing the results of Inter-EC-Bound and Inter-EC, we found that although Inter-EC performs better than Indep, it is far poorer than Inter-EC-Bound

is caused by lots of errors in the predictions of emotion extraction. We can draw the same conclusion when comparing InterCE-Bound and Inter-CE.


具体方法:

Step1: Individual Emotion and Cause Extraction

这部分的目的是为了从文档d中提取可能的情感相关子句原因相关子句,这里分为两种方法:

1. Independent Multi-task Learning
image.png

用白话描述就是对于每个子句c_i,均由词语w_{(i, |c_i|)}组成,是故最下层的BiLSTM-Attention结构用于将子句转为句子级别的表示S_i。然后将得到的表示向量分别给两个BiLSTM进行分类,从而确定Emotion setCause set

  1. Interactive Multi-task Learning

按照作者的表述,InteractiveIndependent的增强版。在这种方法中,也分为以下两种方式:

那么,怎么增强呢?先做其中一个,再把hidden state拿出来做另外一个。

image.png
2. Step 2: Emotion-Cause Pairing and Filtering

这部分的目的是为了对上一步提取出的情感相关子句原因相关子句进行匹配以及筛选。按照以下的步骤来做:

  1. 每两个互相匹配:(c_i^e, c_j^c)
  2. 每对之间计算笛卡尔积:v^d
  3. 构建特征:X_(c_i^e, c_j^c)=[S_i^e, S_j^c, v^d]
  4. Logistic regression model: image.png
上一篇 下一篇

猜你喜欢

热点阅读