论文阅读《TPLinker: Single-stage Joint Extraction of Entities and Relations Through Token Pair Linking》

约定不等于承诺〃 2022-10-19 00:54 138阅读 0赞

# 前言 #

*  文章来源：LawsonAbs（CSDN）
 *  望各位读者审慎阅读

--------------------

# 解码方法 #

之前自己在看论文时，对于这个解码方法没有细致的分析，一带而过。 然后导致老师在课上提问时，我极其尴尬，出来学习，迟早都要补的。今天重新学习一下这个`Decoding algorithm`。

## 为什么要解码？ ##

## 本文怎么解码？ ##

看一下解码算法：

解码步骤如下：

*  step 1:

> For each relation, in the beginning, we extract all entity spans from the EH-to-ET sequence and map each head position to the corresponding entities starting with this position by a dictionary D.

意思就是说：通过一个字典D，要将EH-to-ET中的实体的头位置映射到这个实体。 很奇怪，这里的`dictionary D`是什么意思呢？

*  step 2:

> Next, for each relation,we firstly decode (subject tail position, obect tail postion ) tuples from the ST-to-OT sequence and add them into a set E, and then decode (subject head position, object head position) tuples from the SH-to-OH sequence and lookup all  
> possible entities starting with the head positions in the dictionary D.

也就是说： 集合E中存放的是头实体和尾实体的最末位置的tuple，然后从SH-to-OH序列中解码元组，同时在字典D中查找所有以指定头位置开始的实体。

*  step 3:

> Finally, we iterate all candidate subject-object pairs to check whether their tail positions are in E. If so, a new triplet is extracted and added into the resulting set T.